docs: add comprehensive deployment guide for docker and manual setups

Fix personnel extraction + header bugs on real Polres Cimahi sprint

.claude/settings.local.json

@@ -0,0 +1,18 @@
+{
+  "permissions": {
+    "allow": [
+      "Bash(python -m pytest tests/unit/test_personnel_text_fallback.py -x -q)",
+      "Bash(python -c \"import sys; print\\(sys.executable\\)\")",
+      "Bash(.venv/Scripts/python.exe -m pytest tests/unit/test_personnel_text_fallback.py -x -q)",
+      "Bash(.venv/Scripts/python.exe -m pytest tests/unit -x -q)",
+      "Bash(git stash *)",
+      "Bash(.venv/Scripts/python.exe -m pytest tests/unit/test_api.py::test_documents_sync_returns_pipeline_output -x -q)",
+      "Bash(.venv/Scripts/python.exe -m pytest tests/unit --ignore=tests/unit/test_api.py -q)",
+      "Bash(.venv/Scripts/python.exe -c ' *)",
+      "Bash(xargs grep *)",
+      "Bash(.venv/Scripts/python.exe -m pytest tests/unit -q --ignore=tests/unit/test_api.py --ignore=tests/unit/test_api_hitl.py --ignore=tests/unit/test_blob_storage.py)",
+      "Bash(.venv/Scripts/python.exe -m pytest tests/unit/test_ocr_layout.py tests/unit/test_personnel_text_fallback.py -q)",
+      "Bash(.venv/Scripts/python.exe -m pytest tests/unit/test_personnel_text_fallback.py tests/unit/test_ocr_layout.py -q)"
+    ]
+  }
+}

.env.example

@@ -10,7 +10,8 @@ STORAGE_LOCAL_DIR=./storage
 # ==== OCR ====
 OCR_LANG=latin                # PaddleOCR lang code; "latin" works well for Bahasa Indonesia
 OCR_USE_GPU=false             # set true if running on a GPU host
-OCR_DET_MODEL_DIR=             # leave empty to use PaddleOCR defaults
+# Leave empty to use PaddleOCR defaults.
+OCR_DET_MODEL_DIR=
 OCR_REC_MODEL_DIR=
 OCR_CLS_MODEL_DIR=
 OCR_MAX_IMAGE_SIDE=2200       # downscale longest side before OCR

Makefile

@@ -1,9 +1,10 @@
-.PHONY: help install dev fmt lint typecheck test test-cov run docker-build docker-up docker-down clean
+.PHONY: help install dev update fmt lint typecheck test test-cov run docker-build docker-up docker-down clean
 
 help:
 	@echo "Targets:"
 	@echo "  install       - install runtime + dev deps in current env"
 	@echo "  dev           - run FastAPI app with autoreload"
+	@echo "  update        - git pull + install deps + migrate db + run dev server"
 	@echo "  fmt           - format code with ruff"
 	@echo "  lint          - lint with ruff"
 	@echo "  typecheck     - run mypy"
@@ -21,6 +22,16 @@ install:
 dev:
 	uvicorn ocr_sprint.main:app --reload --host 0.0.0.0 --port 8000
 
+update:
+	@echo "[1/4] Pulling latest code..."
+	git pull
+	@echo "[2/4] Installing/updating dependencies..."
+	pip install -e ".[dev]"
+	@echo "[3/4] Running database migrations..."
+	alembic upgrade head
+	@echo "[4/4] Starting dev server..."
+	uvicorn ocr_sprint.main:app --reload --host 0.0.0.0 --port 8000
+
 fmt:
 	ruff format src tests
 	ruff check --fix src tests

docs/DEPLOYMENT-EXISTING-STACK.md

@@ -0,0 +1,858 @@
+# Deployment OCR Sprint Service (Existing Stack)
+
+Panduan deployment untuk server dengan Python 3.12.3, PostgreSQL 16.13, dan Redis 7.0.15 yang sudah terinstall.
+
+## Informasi Server Anda
+
+- **OS**: Ubuntu 24.04
+- **Python**: 3.12.3 ✅
+- **PostgreSQL**: 16.13 ✅
+- **Redis**: 7.0.15 ✅
+
+Semua versi sudah kompatibel dan optimal untuk OCR Sprint Service!
+
+## Langkah 1: Install System Libraries untuk OpenCV & PaddleOCR
+
+```bash
+# Update package list
+sudo apt update
+
+# Install libraries yang dibutuhkan oleh OpenCV dan PaddleOCR
+sudo apt install -y \
+    libgl1 \
+    libglib2.0-0 \
+    libsm6 \
+    libxext6 \
+    libxrender1 \
+    libgomp1 \
+    libmagic1 \
+    python3.12-venv \
+    python3.12-dev \
+    build-essential \
+    git
+```
+
+## Langkah 2: Setup PostgreSQL Database
+
+```bash
+# Login ke PostgreSQL
+sudo -u postgres psql
+```
+
+Jalankan SQL commands berikut:
+
+```sql
+-- Create user dan database
+CREATE USER ocr WITH PASSWORD '@Offroader123';
+CREATE DATABASE ocr_sprint OWNER ocr;
+
+-- Grant privileges
+GRANT ALL PRIVILEGES ON DATABASE ocr_sprint TO ocr;
+
+-- Connect ke database untuk grant schema privileges
+\c ocr_sprint
+
+-- Grant schema privileges (PostgreSQL 15+)
+GRANT ALL ON SCHEMA public TO ocr;
+GRANT ALL PRIVILEGES ON ALL TABLES IN SCHEMA public TO ocr;
+GRANT ALL PRIVILEGES ON ALL SEQUENCES IN SCHEMA public TO ocr;
+
+-- Verify
+\l ocr_sprint
+\du ocr
+
+-- Exit
+\q
+```
+
+**Generate password yang aman:**
+
+```bash
+# Generate random password
+openssl rand -base64 32
++J33GdYQcWcfqXs169cmgPrQJpLFgybjoedr/tNb0d4=
+```
+
+Simpan password ini, akan digunakan di konfigurasi nanti.
+
+## Langkah 3: Verify Redis
+
+```bash
+# Check Redis status
+sudo systemctl status redis-server
+
+# Test connection
+redis-cli ping
+# Expected output: PONG
+
+# Check Redis config (opsional)
+redis-cli CONFIG GET maxmemory
+```
+
+Jika Redis belum running:
+
+```bash
+sudo systemctl enable redis-server
+sudo systemctl start redis-server
+```
+
+## Langkah 4: Create Application User
+
+```bash
+# Create dedicated user untuk aplikasi
+sudo useradd -m -s /bin/bash ocr
+
+# Create application directory
+sudo mkdir -p /opt/ocr-sprint-service
+sudo chown ocr:ocr /opt/ocr-sprint-service
+```
+
+## Langkah 5: Clone dan Install Application
+
+```bash
+# Switch ke user ocr
+sudo su - ocr
+
+# Clone repository
+cd /opt
+git clone https://github.com/Adriankf59/ocr-sprint-service.git
+cd ocr-sprint-service
+
+# Create virtual environment dengan Python 3.12
+python3.12 -m venv .venv
+
+# Activate virtual environment
+source .venv/bin/activate
+
+# Verify Python version di venv
+python --version
+# Expected: Python 3.12.3
+
+# Upgrade pip
+pip install --upgrade pip setuptools wheel
+
+# Install application dengan OCR dependencies
+# Ini akan download ~1.5GB PaddlePaddle wheels
+pip install -e ".[ocr]"
+
+# Verify installation
+python -c "import paddleocr; print('PaddleOCR OK')"
+python -c "import cv2; print('OpenCV OK')"
+python -c "import fastapi; print('FastAPI OK')"
+```
+
+## Langkah 6: Konfigurasi Application
+
+```bash
+# Masih sebagai user ocr
+cd /opt/ocr-sprint-service
+
+# Copy environment template
+cp .env.example .env
+
+# Edit konfigurasi
+nano .env
+```
+
+**Konfigurasi `/opt/ocr-sprint-service/.env`:**
+
+```bash
+# ==== App ====
+APP_ENV=prod
+APP_HOST=0.0.0.0
+APP_PORT=8000
+APP_LOG_LEVEL=INFO
+
+# ==== Storage ====
+STORAGE_LOCAL_DIR=/opt/ocr-sprint-service/storage
+BLOB_STORAGE_DIR=/opt/ocr-sprint-service/storage/blobs
+BLOB_MAX_UPLOAD_MB=25
+
+# ==== OCR ====
+OCR_LANG=latin
+OCR_USE_GPU=false
+OCR_MAX_IMAGE_SIDE=2200
+
+# ==== Preprocessing ====
+PREPROCESS_TARGET_DPI=300
+PREPROCESS_DENOISE=true
+PREPROCESS_DESKEW=true
+PREPROCESS_DETECT_DOCUMENT=true
+PREPROCESS_REMOVE_SHADOW=true
+PREPROCESS_MIN_QUAD_AREA_FRACTION=0.20
+
+# ==== Table Extraction ====
+TABLES_ENABLED=true
+
+# ==== Confidence ====
+CONFIDENCE_AUTO_APPROVE=0.95
+CONFIDENCE_NEEDS_REVIEW=0.85
+
+# ==== LLM (Phase 5, optional - disable untuk sekarang) ====
+LLM_ENABLED=false
+
+# ==== Async Pipeline ====
+QUEUE_ENABLED=true
+REDIS_URL=redis://localhost:6379/0
+CELERY_TASK_DEFAULT_QUEUE=ocr_sprint
+
+# ==== Database ====
+# Ganti 'your-password-here' dengan password yang Anda generate di Langkah 2
+DATABASE_URL=postgresql+psycopg://ocr:your-password-here@localhost:5432/ocr_sprint
+DATABASE_ECHO=false
+
+# ==== Auth (WAJIB untuk production!) ====
+# Generate dengan: openssl rand -hex 32
+API_KEYS=paste-api-key-1-here,paste-api-key-2-here
+API_KEY_HEADER=X-API-Key
+```
+
+**Generate API keys:**
+
+```bash
+# Generate 2 API keys
+echo "API Key 1: $(openssl rand -hex 32)"
+echo "API Key 2: $(openssl rand -hex 32)"
+```
+
+Copy output dan paste ke `API_KEYS` di file `.env`.
+
+**Create storage directories:**
+
+```bash
+mkdir -p /opt/ocr-sprint-service/storage/blobs
+chmod 755 /opt/ocr-sprint-service/storage
+```
+
+## Langkah 7: Run Database Migrations
+
+```bash
+# Masih sebagai user ocr, dengan venv activated
+cd /opt/ocr-sprint-service
+source .venv/bin/activate
+
+# Run migrations
+alembic upgrade head
+
+# Verify - should show current revision
+alembic current
+
+# Expected output: (head) atau revision number
+```
+
+## Langkah 8: Test Manual Run
+
+```bash
+# Masih sebagai user ocr
+cd /opt/ocr-sprint-service
+source .venv/bin/activate
+
+# Test API server
+uvicorn ocr_sprint.main:app --host 0.0.0.0 --port 8000
+```
+
+**Di terminal lain (sebagai user ubuntu):**
+
+```bash
+# Test health check
+curl http://localhost:8000/api/v1/health
+
+# Expected: {"status":"ok","version":"0.1.0"}
+
+# Test dengan sample file (jika ada)
+curl -X POST "http://localhost:8000/api/v1/documents?sync=true" \
+  -H "X-API-Key: your-api-key-here" \
+  -F "file=@/path/to/test.pdf"
+```
+
+Jika berhasil, stop server dengan `Ctrl+C`.
+
+## Langkah 9: Setup Systemd Services
+
+```bash
+# Exit dari user ocr
+exit
+
+# Kembali sebagai user ubuntu dengan sudo
+```
+
+### Create API Service
+
+```bash
+sudo nano /etc/systemd/system/ocr-sprint-api.service
+```
+
+**Content:**
+
+```ini
+[Unit]
+Description=OCR Sprint API Service
+After=network.target postgresql.service redis-server.service
+Wants=postgresql.service redis-server.service
+
+[Service]
+Type=simple
+User=ocr
+Group=ocr
+WorkingDirectory=/opt/ocr-sprint-service
+
+# Environment
+Environment="PATH=/opt/ocr-sprint-service/.venv/bin:/usr/local/bin:/usr/bin:/bin"
+EnvironmentFile=/opt/ocr-sprint-service/.env
+
+# Start command - 4 workers untuk production
+ExecStart=/opt/ocr-sprint-service/.venv/bin/uvicorn \
+    ocr_sprint.main:app \
+    --host 0.0.0.0 \
+    --port 8000 \
+    --workers 4 \
+    --log-level info
+
+# Restart policy
+Restart=always
+RestartSec=10
+StartLimitInterval=0
+
+# Resource limits
+LimitNOFILE=65536
+
+# Security
+NoNewPrivileges=true
+PrivateTmp=true
+
+[Install]
+WantedBy=multi-user.target
+```
+
+### Create Celery Worker Service
+
+```bash
+sudo nano /etc/systemd/system/ocr-sprint-worker.service
+```
+
+**Content:**
+
+```ini
+[Unit]
+Description=OCR Sprint Celery Worker
+After=network.target postgresql.service redis-server.service ocr-sprint-api.service
+Wants=postgresql.service redis-server.service
+
+[Service]
+Type=simple
+User=ocr
+Group=ocr
+WorkingDirectory=/opt/ocr-sprint-service
+
+# Environment
+Environment="PATH=/opt/ocr-sprint-service/.venv/bin:/usr/local/bin:/usr/bin:/bin"
+EnvironmentFile=/opt/ocr-sprint-service/.env
+
+# Start command - concurrency 2 untuk CPU dengan 4 cores
+# Sesuaikan dengan jumlah CPU cores server Anda
+ExecStart=/opt/ocr-sprint-service/.venv/bin/celery \
+    -A ocr_sprint.worker.celery_app \
+    worker \
+    --loglevel=info \
+    --concurrency=2 \
+    --max-tasks-per-child=100
+
+# Restart policy
+Restart=always
+RestartSec=10
+StartLimitInterval=0
+
+# Resource limits
+LimitNOFILE=65536
+
+# Security
+NoNewPrivileges=true
+PrivateTmp=true
+
+[Install]
+WantedBy=multi-user.target
+```
+
+### Enable dan Start Services
+
+```bash
+# Reload systemd
+sudo systemctl daemon-reload
+
+# Enable services (auto-start on boot)
+sudo systemctl enable ocr-sprint-api
+sudo systemctl enable ocr-sprint-worker
+
+# Start services
+sudo systemctl start ocr-sprint-api
+sudo systemctl start ocr-sprint-worker
+
+# Check status
+sudo systemctl status ocr-sprint-api
+sudo systemctl status ocr-sprint-worker
+```
+
+**Expected output:** `active (running)` dengan warna hijau.
+
+### View Logs
+
+```bash
+# API logs (real-time)
+sudo journalctl -u ocr-sprint-api -f
+
+# Worker logs (real-time)
+sudo journalctl -u ocr-sprint-worker -f
+
+# Last 50 lines
+sudo journalctl -u ocr-sprint-api -n 50
+sudo journalctl -u ocr-sprint-worker -n 50
+```
+
+## Langkah 10: Install dan Setup Nginx
+
+```bash
+# Install Nginx dan Certbot
+sudo apt install -y nginx certbot python3-certbot-nginx
+
+# Check Nginx status
+sudo systemctl status nginx
+```
+
+### Create Nginx Configuration
+
+```bash
+sudo nano /etc/nginx/sites-available/ocr-sprint
+```
+
+**Content (ganti `ocr.yourdomain.com` dengan domain Anda):**
+
+```nginx
+# Upstream
+upstream ocr_api {
+    server 127.0.0.1:8000;
+    keepalive 32;
+}
+
+# Rate limiting
+limit_req_zone $binary_remote_addr zone=api_limit:10m rate=10r/s;
+
+server {
+    listen 80;
+    server_name ocr.yourdomain.com;
+
+    # Max upload size
+    client_max_body_size 30M;
+    client_body_buffer_size 128k;
+
+    # Timeouts
+    proxy_connect_timeout 300s;
+    proxy_send_timeout 300s;
+    proxy_read_timeout 300s;
+    send_timeout 300s;
+
+    # Logging
+    access_log /var/log/nginx/ocr-sprint-access.log;
+    error_log /var/log/nginx/ocr-sprint-error.log;
+
+    # API endpoints
+    location /api/ {
+        limit_req zone=api_limit burst=20 nodelay;
+
+        proxy_pass http://ocr_api;
+        proxy_http_version 1.1;
+        
+        proxy_set_header Host $host;
+        proxy_set_header X-Real-IP $remote_addr;
+        proxy_set_header X-Forwarded-For $proxy_add_x_forwarded_for;
+        proxy_set_header X-Forwarded-Proto $scheme;
+        proxy_set_header Connection "";
+        
+        proxy_buffering off;
+    }
+
+    # Health check
+    location /api/v1/health {
+        proxy_pass http://ocr_api;
+        proxy_http_version 1.1;
+        proxy_set_header Host $host;
+        access_log off;
+    }
+
+    # Metrics (restrict access)
+    location /metrics {
+        allow 127.0.0.1;
+        allow 10.0.0.0/8;
+        deny all;
+
+        proxy_pass http://ocr_api;
+        proxy_http_version 1.1;
+        proxy_set_header Host $host;
+    }
+
+    # API docs
+    location /docs {
+        proxy_pass http://ocr_api;
+        proxy_http_version 1.1;
+        proxy_set_header Host $host;
+    }
+
+    location /redoc {
+        proxy_pass http://ocr_api;
+        proxy_http_version 1.1;
+        proxy_set_header Host $host;
+    }
+}
+```
+
+### Enable Site
+
+```bash
+# Test konfigurasi
+sudo nginx -t
+
+# Enable site
+sudo ln -s /etc/nginx/sites-available/ocr-sprint /etc/nginx/sites-enabled/
+
+# Reload Nginx
+sudo systemctl reload nginx
+```
+
+### Setup SSL (jika punya domain)
+
+```bash
+# Obtain certificate
+sudo certbot --nginx -d ocr.yourdomain.com
+
+# Test auto-renewal
+sudo certbot renew --dry-run
+```
+
+## Langkah 11: Setup Firewall
+
+```bash
+# Check UFW status
+sudo ufw status
+
+# Allow SSH (PENTING!)
+sudo ufw allow 22/tcp
+
+# Allow HTTP dan HTTPS
+sudo ufw allow 80/tcp
+sudo ufw allow 443/tcp
+
+# Enable firewall (jika belum)
+sudo ufw enable
+
+# Verify
+sudo ufw status numbered
+```
+
+## Langkah 12: Verifikasi Final
+
+### Test dari Server
+
+```bash
+# Health check
+curl http://localhost:8000/api/v1/health
+
+# Test async endpoint
+curl -X POST http://localhost:8000/api/v1/documents \
+  -H "X-API-Key: your-api-key-here" \
+  -F "file=@/path/to/test.pdf"
+
+# Expected: {"job_id":"...","status":"pending",...}
+
+# Check job status
+curl -H "X-API-Key: your-api-key-here" \
+  http://localhost:8000/api/v1/documents/JOB_ID_HERE
+```
+
+### Test via Domain (jika sudah setup SSL)
+
+```bash
+curl https://ocr.yourdomain.com/api/v1/health
+```
+
+### Check Services
+
+```bash
+# All services should be active
+sudo systemctl status ocr-sprint-api
+sudo systemctl status ocr-sprint-worker
+sudo systemctl status postgresql
+sudo systemctl status redis-server
+sudo systemctl status nginx
+```
+
+## Monitoring
+
+### View Logs
+
+```bash
+# API logs
+sudo journalctl -u ocr-sprint-api -f
+
+# Worker logs
+sudo journalctl -u ocr-sprint-worker -f
+
+# Nginx access logs
+sudo tail -f /var/log/nginx/ocr-sprint-access.log
+
+# Nginx error logs
+sudo tail -f /var/log/nginx/ocr-sprint-error.log
+```
+
+### Prometheus Metrics
+
+```bash
+# View metrics
+curl http://localhost:8000/metrics
+
+# Key metrics:
+# - ocr_documents_total
+# - ocr_processing_duration_seconds
+# - ocr_confidence_score
+```
+
+## Maintenance
+
+### Restart Services
+
+```bash
+sudo systemctl restart ocr-sprint-api
+sudo systemctl restart ocr-sprint-worker
+```
+
+### Update Application
+
+```bash
+# Switch ke user ocr
+sudo su - ocr
+cd /opt/ocr-sprint-service
+
+# Pull latest code
+git pull
+
+# Activate venv
+source .venv/bin/activate
+
+# Update dependencies
+pip install -e ".[ocr]"
+
+# Run migrations
+alembic upgrade head
+
+# Exit
+exit
+
+# Restart services
+sudo systemctl restart ocr-sprint-api
+sudo systemctl restart ocr-sprint-worker
+
+# Check logs
+sudo journalctl -u ocr-sprint-api -n 50
+```
+
+### Database Backup
+
+```bash
+# Create backup directory
+sudo mkdir -p /opt/ocr-sprint-service/backups
+sudo chown ocr:ocr /opt/ocr-sprint-service/backups
+
+# Manual backup
+sudo -u ocr pg_dump -h localhost -U ocr ocr_sprint | gzip > /opt/ocr-sprint-service/backups/backup_$(date +%Y%m%d_%H%M%S).sql.gz
+```
+
+**Setup automated backup:**
+
+```bash
+# Create backup script
+sudo nano /opt/ocr-sprint-service/backup.sh
+```
+
+```bash
+#!/bin/bash
+BACKUP_DIR="/opt/ocr-sprint-service/backups"
+DATE=$(date +%Y%m%d_%H%M%S)
+
+mkdir -p $BACKUP_DIR
+
+# Backup database
+PGPASSWORD='your-db-password' pg_dump -h localhost -U ocr ocr_sprint | gzip > $BACKUP_DIR/db_$DATE.sql.gz
+
+# Keep only last 7 days
+find $BACKUP_DIR -name "db_*.sql.gz" -mtime +7 -delete
+
+echo "Backup completed: $DATE"
+```
+
+```bash
+# Make executable
+sudo chmod +x /opt/ocr-sprint-service/backup.sh
+sudo chown ocr:ocr /opt/ocr-sprint-service/backup.sh
+
+# Setup cron (daily at 2 AM)
+sudo crontab -e -u ocr
+
+# Add line:
+0 2 * * * /opt/ocr-sprint-service/backup.sh >> /var/log/ocr-backup.log 2>&1
+```
+
+## Troubleshooting
+
+### Service tidak start
+
+```bash
+# Check detailed logs
+sudo journalctl -u ocr-sprint-api -n 100 --no-pager
+sudo journalctl -u ocr-sprint-worker -n 100 --no-pager
+
+# Check file permissions
+ls -la /opt/ocr-sprint-service
+ls -la /opt/ocr-sprint-service/storage
+
+# Test manual run
+sudo su - ocr
+cd /opt/ocr-sprint-service
+source .venv/bin/activate
+uvicorn ocr_sprint.main:app --host 0.0.0.0 --port 8000
+```
+
+### Database connection error
+
+```bash
+# Test connection
+sudo -u ocr psql -h localhost -U ocr -d ocr_sprint
+
+# Check PostgreSQL status
+sudo systemctl status postgresql
+
+# Check PostgreSQL logs
+sudo journalctl -u postgresql -n 50
+```
+
+### Redis connection error
+
+```bash
+# Test Redis
+redis-cli ping
+
+# Check Redis status
+sudo systemctl status redis-server
+
+# Check Redis logs
+sudo journalctl -u redis-server -n 50
+```
+
+### Worker tidak memproses jobs
+
+```bash
+# Check Celery worker status
+sudo su - ocr
+cd /opt/ocr-sprint-service
+source .venv/bin/activate
+celery -A ocr_sprint.worker.celery_app inspect active
+celery -A ocr_sprint.worker.celery_app inspect stats
+
+# Check Redis queue
+redis-cli LLEN ocr_sprint
+```
+
+### PaddleOCR error
+
+```bash
+# Re-download models
+sudo su - ocr
+cd /opt/ocr-sprint-service
+source .venv/bin/activate
+
+python << EOF
+from paddleocr import PaddleOCR
+ocr = PaddleOCR(use_angle_cls=True, lang='latin')
+print("Models downloaded successfully")
+EOF
+```
+
+## Performance Tuning
+
+### Check CPU cores
+
+```bash
+nproc
+```
+
+### Adjust worker concurrency
+
+```bash
+# Edit worker service
+sudo nano /etc/systemd/system/ocr-sprint-worker.service
+
+# Untuk 4 cores: --concurrency=2
+# Untuk 8 cores: --concurrency=4
+# Untuk 16 cores: --concurrency=8
+
+# Reload dan restart
+sudo systemctl daemon-reload
+sudo systemctl restart ocr-sprint-worker
+```
+
+### PostgreSQL 16 Tuning
+
+```bash
+sudo nano /etc/postgresql/16/main/postgresql.conf
+```
+
+**Recommended settings (sesuaikan dengan RAM server):**
+
+```
+# Untuk 8GB RAM:
+shared_buffers = 2GB
+effective_cache_size = 6GB
+maintenance_work_mem = 512MB
+work_mem = 8MB
+
+# Untuk 16GB RAM:
+shared_buffers = 4GB
+effective_cache_size = 12GB
+maintenance_work_mem = 1GB
+work_mem = 10MB
+
+# General
+checkpoint_completion_target = 0.9
+wal_buffers = 16MB
+default_statistics_target = 100
+random_page_cost = 1.1
+effective_io_concurrency = 200
+max_worker_processes = 4
+max_parallel_workers_per_gather = 2
+max_parallel_workers = 4
+```
+
+```bash
+sudo systemctl restart postgresql
+```
+
+## Security Checklist
+
+- [ ] API keys set dengan nilai random yang kuat
+- [ ] Database password diganti dari default
+- [ ] Firewall enabled (UFW)
+- [ ] SSL/TLS enabled (jika punya domain)
+- [ ] `/metrics` endpoint restricted
+- [ ] PostgreSQL hanya listen di localhost
+- [ ] Redis hanya listen di localhost
+- [ ] Backup automated (cron job)
+- [ ] OS security updates enabled
+
+## Next Steps
+
+1. **Setup monitoring** - Install Prometheus + Grafana (opsional)
+2. **Setup alerting** - Email/Slack notification untuk errors
+3. **Load testing** - Test dengan volume dokumen production
+4. **Backup verification** - Test restore dari backup
+5. **Documentation** - Dokumentasi API keys untuk tim
+
+## Support
+
+Untuk pertanyaan atau issues, hubungi tim development.

docs/DEPLOYMENT-GUIDE.md

@@ -0,0 +1,571 @@
+# Panduan Deployment OCR Sprint Service
+
+> Dokumen ini adalah panduan langkah-langkah deployment **ocr-sprint-service** ke server production. Disusun berdasarkan kondisi kodingan aktual per April 2026 (Phase 1–4 selesai).
+
+---
+
+## Daftar Isi
+
+1. [Gambaran Arsitektur](#1-gambaran-arsitektur)
+2. [Prasyarat Server](#2-prasyarat-server)
+3. [Opsi A — Docker Compose (Recommended)](#3-opsi-a--docker-compose-recommended)
+4. [Opsi B — Manual (Tanpa Docker)](#4-opsi-b--manual-tanpa-docker)
+5. [Konfigurasi Environment Production](#5-konfigurasi-environment-production)
+6. [Reverse Proxy & SSL (Nginx)](#6-reverse-proxy--ssl-nginx)
+7. [Firewall](#7-firewall)
+8. [Verifikasi Deployment](#8-verifikasi-deployment)
+9. [Monitoring & Maintenance](#9-monitoring--maintenance)
+10. [Troubleshooting](#10-troubleshooting)
+11. [Security Checklist](#11-security-checklist)
+
+---
+
+## 1. Gambaran Arsitektur
+
+```
+┌──────────┐     ┌──────────────┐     ┌───────┐
+│  Client  │────▶│  Nginx (SSL) │────▶│  API  │──▶ PaddleOCR
+└──────────┘     └──────────────┘     │ :8000 │      Pipeline
+                                      └───┬───┘
+                                          │ async job
+                                    ┌─────▼─────┐
+                                    │   Redis    │
+                                    │   :6379    │
+                                    └─────┬─────┘
+                                    ┌─────▼──────┐
+                                    │   Worker   │──▶ PaddleOCR
+                                    │  (Celery)  │      Pipeline
+                                    └─────┬──────┘
+                                    ┌─────▼──────┐
+                                    │ PostgreSQL │
+                                    │   :5432    │
+                                    └────────────┘
+```
+
+**4 services** yang harus berjalan:
+
+| Service | Fungsi |
+|---------|--------|
+| **API** (FastAPI + Uvicorn) | Menerima upload dokumen, serve hasil OCR |
+| **Worker** (Celery) | Async OCR processing di background |
+| **Redis** | Message broker untuk job queue |
+| **PostgreSQL** | Menyimpan job state & hasil ekstraksi |
+
+Blob storage menggunakan **local filesystem** (belum S3/MinIO).
+
+---
+
+## 2. Prasyarat Server
+
+### Spesifikasi Minimum
+
+| Resource | Minimum | Recommended |
+|----------|---------|-------------|
+| OS | Ubuntu 20.04+ / Debian 11+ | Ubuntu 22.04+ |
+| CPU | 4 cores | 8 cores |
+| RAM | 8 GB | 16 GB |
+| Storage | 50 GB free | 100 GB free |
+| Python | 3.10–3.12 | 3.11 atau 3.12 |
+| Network | Port 8000 (internal) | + Port 80/443 (Nginx) |
+
+### Kebutuhan Disk
+
+- ~1.5 GB — PaddlePaddle wheels
+- ~200 MB — PaddleOCR model downloads (otomatis saat pertama jalan)
+- Sisanya — blob storage dokumen yang diupload
+
+### Software yang Dibutuhkan
+
+- **Docker Compose** — untuk Opsi A
+- **Python 3.10–3.12 + PostgreSQL + Redis** — untuk Opsi B
+- **Git** — kedua opsi
+- **Nginx** (opsional) — reverse proxy + SSL
+
+---
+
+## 3. Opsi A — Docker Compose (Recommended)
+
+> Cara paling cepat. Semua service (API, Worker, Redis, Postgres) berjalan dalam container.
+
+### 3.1 Login & Clone
+
+```bash
+ssh user@your-server.com
+
+git clone https://github.com/Adriankf59/ocr-sprint-service.git
+cd ocr-sprint-service
+```
+
+### 3.2 Konfigurasi .env
+
+```bash
+cp .env.example .env
+nano .env
+```
+
+Lihat [Bagian 5](#5-konfigurasi-environment-production) untuk detail konfigurasi production.
+
+> [!IMPORTANT]
+> Untuk Docker Compose, **jangan ubah** `DATABASE_URL` dan `REDIS_URL` — sudah dioverride oleh `docker-compose.yml` via environment variables di masing-masing container.
+
+### 3.3 Build & Start
+
+```bash
+# Build image (~5–10 menit pertama kali)
+docker compose build
+
+# Start semua services
+docker compose up -d
+
+# Cek logs
+docker compose logs -f api worker
+```
+
+Container `api` akan otomatis menjalankan `alembic upgrade head` sebelum start server (lihat `command` di `docker-compose.yml`).
+
+### 3.4 First-Run Model Download
+
+Request pertama akan trigger download model PaddleOCR (~200 MB) ke Docker volume `paddle-models`. Tunggu hingga selesai sebelum test.
+
+```bash
+# Monitor download di logs
+docker compose logs -f api
+```
+
+### 3.5 Verifikasi
+
+```bash
+curl http://localhost:8000/api/v1/health
+# Expected: {"status":"ok","version":"0.1.0"}
+```
+
+### 3.6 Update Service (Setelah Ada Perubahan Kode)
+
+```bash
+cd ocr-sprint-service
+git pull
+docker compose build
+docker compose up -d
+```
+
+---
+
+## 4. Opsi B — Manual (Tanpa Docker)
+
+> Untuk server yang sudah punya Python, PostgreSQL, dan Redis terinstall.
+
+### 4.1 Install System Libraries
+
+```bash
+sudo apt update && sudo apt upgrade -y
+
+# Libraries untuk OpenCV & PaddleOCR
+sudo apt install -y \
+    python3.11 python3.11-venv python3.11-dev \
+    libgl1 libglib2.0-0 libsm6 libxext6 libxrender1 \
+    libgomp1 libmagic1 \
+    build-essential git curl
+
+# Install Redis & PostgreSQL (jika belum ada)
+sudo apt install -y redis-server postgresql postgresql-contrib
+sudo systemctl enable --now redis-server postgresql
+```
+
+> [!NOTE]
+> Jika server sudah punya Python 3.12, gunakan `python3.12` di semua perintah selanjutnya.
+
+### 4.2 Setup Database
+
+```bash
+sudo -u postgres psql
+```
+
+```sql
+CREATE USER ocr WITH PASSWORD 'ganti-password-kuat';
+CREATE DATABASE ocr_sprint OWNER ocr;
+GRANT ALL PRIVILEGES ON DATABASE ocr_sprint TO ocr;
+\c ocr_sprint
+GRANT ALL ON SCHEMA public TO ocr;
+\q
+```
+
+### 4.3 Create Application User & Directory
+
+```bash
+sudo useradd -m -s /bin/bash ocr
+sudo mkdir -p /opt/ocr-sprint-service
+sudo chown ocr:ocr /opt/ocr-sprint-service
+```
+
+### 4.4 Clone & Install
+
+```bash
+sudo su - ocr
+cd /opt
+git clone https://github.com/Adriankf59/ocr-sprint-service.git
+cd ocr-sprint-service
+
+# Create virtual environment
+python3.11 -m venv .venv
+source .venv/bin/activate
+
+# Install dependencies + OCR runtime (~1.5 GB download)
+pip install --upgrade pip setuptools wheel
+pip install -e ".[ocr]"
+
+# Verify
+python -c "import paddleocr; print('PaddleOCR OK')"
+python -c "import fastapi; print('FastAPI OK')"
+```
+
+### 4.5 Konfigurasi .env
+
+```bash
+cp .env.example .env
+nano .env
+```
+
+**Wajib diubah untuk manual deployment:**
+
+```bash
+APP_ENV=prod
+DATABASE_URL=postgresql+psycopg://ocr:ganti-password-kuat@localhost:5432/ocr_sprint
+REDIS_URL=redis://localhost:6379/0
+QUEUE_ENABLED=true
+API_KEYS=your-generated-api-key
+STORAGE_LOCAL_DIR=/opt/ocr-sprint-service/storage
+BLOB_STORAGE_DIR=/opt/ocr-sprint-service/storage/blobs
+```
+
+```bash
+# Create storage directories
+mkdir -p /opt/ocr-sprint-service/storage/blobs
+```
+
+### 4.6 Run Database Migrations
+
+```bash
+source .venv/bin/activate
+alembic upgrade head
+alembic current  # verify
+```
+
+### 4.7 Test Manual
+
+```bash
+uvicorn ocr_sprint.main:app --host 0.0.0.0 --port 8000
+# Di terminal lain: curl http://localhost:8000/api/v1/health
+# Ctrl+C untuk stop
+```
+
+### 4.8 Setup Systemd Services
+
+**API Service** — `/etc/systemd/system/ocr-sprint-api.service`:
+
+```ini
+[Unit]
+Description=OCR Sprint API Service
+After=network.target postgresql.service redis-server.service
+
+[Service]
+Type=simple
+User=ocr
+Group=ocr
+WorkingDirectory=/opt/ocr-sprint-service
+Environment="PATH=/opt/ocr-sprint-service/.venv/bin:/usr/local/bin:/usr/bin:/bin"
+EnvironmentFile=/opt/ocr-sprint-service/.env
+ExecStart=/opt/ocr-sprint-service/.venv/bin/uvicorn \
+    ocr_sprint.main:app \
+    --host 0.0.0.0 --port 8000 --workers 4 --log-level info
+Restart=always
+RestartSec=10
+LimitNOFILE=65536
+NoNewPrivileges=true
+
+[Install]
+WantedBy=multi-user.target
+```
+
+**Worker Service** — `/etc/systemd/system/ocr-sprint-worker.service`:
+
+```ini
+[Unit]
+Description=OCR Sprint Celery Worker
+After=network.target postgresql.service redis-server.service
+
+[Service]
+Type=simple
+User=ocr
+Group=ocr
+WorkingDirectory=/opt/ocr-sprint-service
+Environment="PATH=/opt/ocr-sprint-service/.venv/bin:/usr/local/bin:/usr/bin:/bin"
+EnvironmentFile=/opt/ocr-sprint-service/.env
+ExecStart=/opt/ocr-sprint-service/.venv/bin/celery \
+    -A ocr_sprint.worker.celery_app worker \
+    --loglevel=info --concurrency=2 --max-tasks-per-child=100
+Restart=always
+RestartSec=10
+LimitNOFILE=65536
+NoNewPrivileges=true
+
+[Install]
+WantedBy=multi-user.target
+```
+
+**Enable & Start:**
+
+```bash
+# Keluar dari user ocr dulu
+exit
+
+sudo systemctl daemon-reload
+sudo systemctl enable --now ocr-sprint-api ocr-sprint-worker
+sudo systemctl status ocr-sprint-api ocr-sprint-worker
+```
+
+### 4.9 Update Service (Manual)
+
+```bash
+sudo su - ocr
+cd /opt/ocr-sprint-service
+git pull
+source .venv/bin/activate
+pip install -e ".[ocr]"
+alembic upgrade head
+exit
+
+sudo systemctl restart ocr-sprint-api ocr-sprint-worker
+```
+
+---
+
+## 5. Konfigurasi Environment Production
+
+Berikut konfigurasi `.env` yang **wajib diubah** dari default untuk production:
+
+| Variable | Default | Production | Keterangan |
+|----------|---------|------------|------------|
+| `APP_ENV` | `local` | `prod` | Mode environment |
+| `API_KEYS` | *(kosong)* | `key1,key2` | **WAJIB!** Auth disabled jika kosong |
+| `QUEUE_ENABLED` | `false` | `true` | Aktifkan async processing |
+| `DATABASE_URL` | `sqlite:///...` | `postgresql+psycopg://...` | Docker: otomatis di-override |
+| `REDIS_URL` | `redis://localhost:6379/0` | Sesuaikan | Docker: otomatis di-override |
+| `OCR_USE_GPU` | `false` | `true` jika ada GPU | Mode GPU butuh NVIDIA driver |
+| `TABLES_ENABLED` | `true` | `true` | Ekstraksi tabel personel |
+
+**Generate API Key:**
+
+```bash
+openssl rand -hex 32
+```
+
+> [!WARNING]
+> Jangan pernah deploy ke production tanpa mengisi `API_KEYS`. Jika kosong, semua endpoint terbuka tanpa autentikasi.
+
+---
+
+## 6. Reverse Proxy & SSL (Nginx)
+
+### Install
+
+```bash
+sudo apt install -y nginx certbot python3-certbot-nginx
+```
+
+### Konfigurasi — `/etc/nginx/sites-available/ocr-sprint`
+
+```nginx
+upstream ocr_api {
+    server 127.0.0.1:8000;
+    keepalive 32;
+}
+
+server {
+    listen 80;
+    server_name ocr.yourdomain.com;
+
+    client_max_body_size 30M;
+
+    proxy_connect_timeout 300s;
+    proxy_read_timeout 300s;
+
+    location / {
+        proxy_pass http://ocr_api;
+        proxy_http_version 1.1;
+        proxy_set_header Host $host;
+        proxy_set_header X-Real-IP $remote_addr;
+        proxy_set_header X-Forwarded-For $proxy_add_x_forwarded_for;
+        proxy_set_header X-Forwarded-Proto $scheme;
+    }
+
+    location /metrics {
+        allow 127.0.0.1;
+        allow 10.0.0.0/8;
+        deny all;
+        proxy_pass http://ocr_api;
+    }
+}
+```
+
+### Enable & SSL
+
+```bash
+sudo ln -s /etc/nginx/sites-available/ocr-sprint /etc/nginx/sites-enabled/
+sudo nginx -t
+sudo systemctl reload nginx
+
+# SSL
+sudo certbot --nginx -d ocr.yourdomain.com
+```
+
+---
+
+## 7. Firewall
+
+```bash
+sudo ufw allow 22/tcp    # SSH — PENTING!
+sudo ufw allow 80/tcp    # HTTP
+sudo ufw allow 443/tcp   # HTTPS
+sudo ufw enable
+sudo ufw status
+```
+
+> [!CAUTION]
+> Pastikan SSH (port 22) di-allow **sebelum** enable firewall, agar tidak terkunci dari server.
+
+---
+
+## 8. Verifikasi Deployment
+
+### Health Check
+
+```bash
+curl http://localhost:8000/api/v1/health
+# {"status":"ok","version":"0.1.0"}
+```
+
+### Test OCR (Sync)
+
+```bash
+curl -X POST "http://localhost:8000/api/v1/documents?sync=true" \
+  -H "X-API-Key: your-api-key" \
+  -F "file=@/path/to/test.pdf" | jq
+```
+
+### Test OCR (Async — Production Flow)
+
+```bash
+# Submit job
+curl -X POST http://localhost:8000/api/v1/documents \
+  -H "X-API-Key: your-api-key" \
+  -F "file=@document.pdf" | jq
+# → {"job_id":"8f2a...","status":"pending",...}
+
+# Poll result
+curl -H "X-API-Key: your-api-key" \
+  http://localhost:8000/api/v1/documents/8f2a... | jq
+# → {"status":"completed","confidence":0.93,"data":{...}}
+```
+
+### Cek Semua Service Berjalan
+
+```bash
+# Docker
+docker compose ps
+
+# Manual
+sudo systemctl status ocr-sprint-api ocr-sprint-worker postgresql redis-server nginx
+```
+
+---
+
+## 9. Monitoring & Maintenance
+
+### Logs
+
+```bash
+# Docker
+docker compose logs -f api worker
+
+# Manual (systemd)
+sudo journalctl -u ocr-sprint-api -f
+sudo journalctl -u ocr-sprint-worker -f
+```
+
+### Prometheus Metrics
+
+```bash
+curl http://localhost:8000/metrics
+```
+
+Metrics penting: `ocr_documents_total`, `ocr_processing_duration_seconds`, `ocr_confidence_score`.
+
+### Backup Database
+
+```bash
+# Docker
+docker compose exec postgres pg_dump -U ocr ocr_sprint > backup_$(date +%Y%m%d).sql
+
+# Manual
+pg_dump -U ocr -h localhost ocr_sprint | gzip > backup_$(date +%Y%m%d).sql.gz
+```
+
+### Automated Backup (Cron)
+
+```bash
+# /opt/ocr-sprint-service/backup.sh
+#!/bin/bash
+BACKUP_DIR="/opt/ocr-sprint-service/backups"
+mkdir -p $BACKUP_DIR
+pg_dump -U ocr -h localhost ocr_sprint | gzip > $BACKUP_DIR/db_$(date +%Y%m%d_%H%M%S).sql.gz
+find $BACKUP_DIR -name "db_*.sql.gz" -mtime +7 -delete
+```
+
+```bash
+chmod +x /opt/ocr-sprint-service/backup.sh
+# Cron: daily at 2 AM
+echo "0 2 * * * /opt/ocr-sprint-service/backup.sh >> /var/log/ocr-backup.log 2>&1" | sudo crontab -u ocr -
+```
+
+---
+
+## 10. Troubleshooting
+
+| Masalah | Diagnosis | Solusi |
+|---------|-----------|--------|
+| Service tidak start | `journalctl -u ocr-sprint-api -n 100` | Cek permissions, `.env`, dan log error |
+| PaddleOCR model gagal download | Timeout di logs | `python -c "from paddleocr import PaddleOCR; PaddleOCR(lang='latin')"` |
+| Worker tidak proses jobs | `redis-cli ping` → bukan PONG | Pastikan Redis running, cek `REDIS_URL` |
+| Database migration error | `alembic current` | `alembic stamp head` lalu `alembic upgrade head` |
+| Port 8000 sudah dipakai | `ss -tlnp | grep 8000` | Kill proses lama atau ganti port di `.env` |
+| Out of memory | OOM killer di logs | Kurangi `--concurrency` di worker, atau tambah RAM |
+
+---
+
+## 11. Security Checklist
+
+- [ ] `API_KEYS` diisi dengan random key (`openssl rand -hex 32`)
+- [ ] Password database diganti dari default
+- [ ] Firewall aktif (hanya port 22, 80, 443 terbuka)
+- [ ] SSL/TLS aktif via Nginx + Let's Encrypt
+- [ ] Endpoint `/metrics` restricted ke internal network
+- [ ] Backup database otomatis via cron
+- [ ] OS security updates enabled (`unattended-upgrades`)
+- [ ] `APP_ENV=prod` (bukan `local`)
+
+---
+
+## Quick Reference — Perintah Sehari-hari
+
+```bash
+# === Docker ===
+docker compose up -d          # Start
+docker compose down            # Stop
+docker compose logs -f api     # Logs
+docker compose build && docker compose up -d  # Update
+
+# === Manual ===
+sudo systemctl restart ocr-sprint-api ocr-sprint-worker  # Restart
+sudo journalctl -u ocr-sprint-api -f                     # Logs
+curl http://localhost:8000/api/v1/health                  # Health check
+```

docs/DEPLOYMENT-MANUAL.md

@@ -0,0 +1,943 @@
+# Deployment Manual OCR Sprint Service (Tanpa Docker)
+
+Panduan lengkap deployment OCR Sprint Service langsung di server tanpa menggunakan Docker.
+
+## Prasyarat Server
+
+### Spesifikasi Minimum
+- **OS**: Ubuntu 20.04+ / Debian 11+ / RHEL 8+
+- **CPU**: 4 cores (8 cores recommended)
+- **RAM**: 8 GB minimum (16 GB recommended)
+- **Storage**: 50 GB free space
+- **User**: Non-root user dengan sudo access
+
+### Port yang Dibutuhkan
+- `8000`: API server (internal, akan di-proxy oleh Nginx)
+- `80/443`: HTTP/HTTPS (Nginx)
+- `5432`: PostgreSQL (localhost only)
+- `6379`: Redis (localhost only)
+
+## Langkah 1: Install System Dependencies
+
+### Ubuntu/Debian
+
+```bash
+# Update system
+sudo apt update && sudo apt upgrade -y
+
+# Install Python 3.11
+sudo apt install -y software-properties-common
+sudo add-apt-repository ppa:deadsnakes/ppa -y
+sudo apt update
+sudo apt install -y python3.11 python3.11-venv python3.11-dev python3-pip
+
+# Install system libraries untuk OpenCV dan PaddleOCR
+sudo apt install -y \
+    libgl1-mesa-glx \
+    libglib2.0-0 \
+    libsm6 \
+    libxext6 \
+    libxrender1 \
+    libgomp1 \
+    libmagic1 \
+    build-essential \
+    git \
+    curl \
+    wget
+
+# Install Redis
+sudo apt install -y redis-server
+sudo systemctl enable redis-server
+sudo systemctl start redis-server
+
+# Install PostgreSQL
+sudo apt install -y postgresql postgresql-contrib
+sudo systemctl enable postgresql
+sudo systemctl start postgresql
+```
+
+### RHEL/CentOS/Rocky Linux
+
+```bash
+# Update system
+sudo dnf update -y
+
+# Install Python 3.11
+sudo dnf install -y python3.11 python3.11-devel python3.11-pip
+
+# Install system libraries
+sudo dnf install -y \
+    mesa-libGL \
+    glib2 \
+    libSM \
+    libXext \
+    libXrender \
+    file-libs \
+    gcc \
+    gcc-c++ \
+    make \
+    git
+
+# Install Redis
+sudo dnf install -y redis
+sudo systemctl enable redis
+sudo systemctl start redis
+
+# Install PostgreSQL
+sudo dnf install -y postgresql-server postgresql-contrib
+sudo postgresql-setup --initdb
+sudo systemctl enable postgresql
+sudo systemctl start postgresql
+```
+
+## Langkah 2: Setup Database PostgreSQL
+
+```bash
+# Masuk sebagai postgres user
+sudo -u postgres psql
+
+# Jalankan SQL commands berikut:
+```
+
+```sql
+-- Create user dan database
+CREATE USER ocr WITH PASSWORD 'ganti-dengan-password-kuat';
+CREATE DATABASE ocr_sprint OWNER ocr;
+
+-- Grant privileges
+GRANT ALL PRIVILEGES ON DATABASE ocr_sprint TO ocr;
+
+-- Connect ke database
+\c ocr_sprint
+
+-- Grant schema privileges (PostgreSQL 15+)
+GRANT ALL ON SCHEMA public TO ocr;
+
+-- Exit
+\q
+```
+
+**Konfigurasi PostgreSQL untuk remote access (opsional):**
+
+```bash
+# Edit postgresql.conf
+sudo nano /etc/postgresql/14/main/postgresql.conf
+
+# Uncomment dan ubah:
+listen_addresses = 'localhost'  # Tetap localhost untuk keamanan
+
+# Edit pg_hba.conf
+sudo nano /etc/postgresql/14/main/pg_hba.conf
+
+# Tambahkan line:
+local   ocr_sprint      ocr                                     scram-sha-256
+
+# Restart PostgreSQL
+sudo systemctl restart postgresql
+```
+
+## Langkah 3: Setup Application User
+
+```bash
+# Create dedicated user untuk aplikasi
+sudo useradd -m -s /bin/bash ocr
+sudo usermod -aG sudo ocr  # Opsional, untuk maintenance
+
+# Create application directory
+sudo mkdir -p /opt/ocr-sprint-service
+sudo chown ocr:ocr /opt/ocr-sprint-service
+
+# Switch ke user ocr
+sudo su - ocr
+```
+
+## Langkah 4: Install Application
+
+```bash
+# Clone repository
+cd /opt
+git clone https://github.com/Adriankf59/ocr-sprint-service.git
+cd ocr-sprint-service
+
+# Create virtual environment
+python3.11 -m venv .venv
+
+# Activate virtual environment
+source .venv/bin/activate
+
+# Upgrade pip
+pip install --upgrade pip setuptools wheel
+
+# Install application dengan OCR dependencies
+pip install -e ".[ocr]"
+
+# Verify installation
+python -c "import paddleocr; print('PaddleOCR installed successfully')"
+```
+
+## Langkah 5: Konfigurasi Application
+
+```bash
+# Copy environment template
+cp .env.example .env
+
+# Edit konfigurasi
+nano .env
+```
+
+**Konfigurasi production (`/opt/ocr-sprint-service/.env`):**
+
+```bash
+# ==== App ====
+APP_ENV=prod
+APP_HOST=0.0.0.0
+APP_PORT=8000
+APP_LOG_LEVEL=INFO
+
+# ==== Storage ====
+STORAGE_LOCAL_DIR=/opt/ocr-sprint-service/storage
+BLOB_STORAGE_DIR=/opt/ocr-sprint-service/storage/blobs
+BLOB_MAX_UPLOAD_MB=25
+
+# ==== OCR ====
+OCR_LANG=latin
+OCR_USE_GPU=false
+OCR_MAX_IMAGE_SIDE=2200
+
+# ==== Preprocessing ====
+PREPROCESS_TARGET_DPI=300
+PREPROCESS_DENOISE=true
+PREPROCESS_DESKEW=true
+PREPROCESS_DETECT_DOCUMENT=true
+PREPROCESS_REMOVE_SHADOW=true
+PREPROCESS_MIN_QUAD_AREA_FRACTION=0.20
+
+# ==== Table Extraction ====
+TABLES_ENABLED=true
+
+# ==== Confidence ====
+CONFIDENCE_AUTO_APPROVE=0.95
+CONFIDENCE_NEEDS_REVIEW=0.85
+
+# ==== LLM (Phase 5, optional) ====
+LLM_ENABLED=false
+
+# ==== Async Pipeline ====
+QUEUE_ENABLED=true
+REDIS_URL=redis://localhost:6379/0
+CELERY_TASK_DEFAULT_QUEUE=ocr_sprint
+
+# ==== Database ====
+DATABASE_URL=postgresql+psycopg://ocr:ganti-dengan-password-kuat@localhost:5432/ocr_sprint
+DATABASE_ECHO=false
+
+# ==== Auth (WAJIB!) ====
+API_KEYS=key1-ganti-dengan-random-string,key2-ganti-dengan-random-string
+API_KEY_HEADER=X-API-Key
+```
+
+**Generate secure API keys:**
+
+```bash
+# Generate 2 API keys
+openssl rand -hex 32
+openssl rand -hex 32
+```
+
+**Create storage directories:**
+
+```bash
+mkdir -p /opt/ocr-sprint-service/storage/blobs
+chmod 755 /opt/ocr-sprint-service/storage
+```
+
+## Langkah 6: Run Database Migrations
+
+```bash
+# Masih sebagai user ocr, dengan venv activated
+cd /opt/ocr-sprint-service
+source .venv/bin/activate
+
+# Run migrations
+alembic upgrade head
+
+# Verify
+alembic current
+```
+
+## Langkah 7: Test Manual Run
+
+```bash
+# Test API server
+uvicorn ocr_sprint.main:app --host 0.0.0.0 --port 8000
+
+# Di terminal lain, test health check
+curl http://localhost:8000/api/v1/health
+
+# Jika berhasil, stop dengan Ctrl+C
+```
+
+## Langkah 8: Setup Systemd Services
+
+### API Service
+
+```bash
+# Exit dari user ocr, kembali ke user dengan sudo
+exit
+
+# Create systemd service file
+sudo nano /etc/systemd/system/ocr-sprint-api.service
+```
+
+**Content `/etc/systemd/system/ocr-sprint-api.service`:**
+
+```ini
+[Unit]
+Description=OCR Sprint API Service
+After=network.target postgresql.service redis.service
+Wants=postgresql.service redis.service
+
+[Service]
+Type=simple
+User=ocr
+Group=ocr
+WorkingDirectory=/opt/ocr-sprint-service
+
+# Environment
+Environment="PATH=/opt/ocr-sprint-service/.venv/bin:/usr/local/bin:/usr/bin:/bin"
+EnvironmentFile=/opt/ocr-sprint-service/.env
+
+# Start command - 4 workers untuk production
+ExecStart=/opt/ocr-sprint-service/.venv/bin/uvicorn \
+    ocr_sprint.main:app \
+    --host 0.0.0.0 \
+    --port 8000 \
+    --workers 4 \
+    --log-level info
+
+# Restart policy
+Restart=always
+RestartSec=10
+StartLimitInterval=0
+
+# Resource limits
+LimitNOFILE=65536
+MemoryLimit=6G
+
+# Security
+NoNewPrivileges=true
+PrivateTmp=true
+
+[Install]
+WantedBy=multi-user.target
+```
+
+### Celery Worker Service
+
+```bash
+sudo nano /etc/systemd/system/ocr-sprint-worker.service
+```
+
+**Content `/etc/systemd/system/ocr-sprint-worker.service`:**
+
+```ini
+[Unit]
+Description=OCR Sprint Celery Worker
+After=network.target postgresql.service redis.service ocr-sprint-api.service
+Wants=postgresql.service redis.service
+
+[Service]
+Type=simple
+User=ocr
+Group=ocr
+WorkingDirectory=/opt/ocr-sprint-service
+
+# Environment
+Environment="PATH=/opt/ocr-sprint-service/.venv/bin:/usr/local/bin:/usr/bin:/bin"
+EnvironmentFile=/opt/ocr-sprint-service/.env
+
+# Start command - concurrency 2 untuk 4 core CPU
+ExecStart=/opt/ocr-sprint-service/.venv/bin/celery \
+    -A ocr_sprint.worker.celery_app \
+    worker \
+    --loglevel=info \
+    --concurrency=2 \
+    --max-tasks-per-child=100
+
+# Restart policy
+Restart=always
+RestartSec=10
+StartLimitInterval=0
+
+# Resource limits
+LimitNOFILE=65536
+MemoryLimit=4G
+
+# Security
+NoNewPrivileges=true
+PrivateTmp=true
+
+[Install]
+WantedBy=multi-user.target
+```
+
+### Enable dan Start Services
+
+```bash
+# Reload systemd
+sudo systemctl daemon-reload
+
+# Enable services (auto-start on boot)
+sudo systemctl enable ocr-sprint-api
+sudo systemctl enable ocr-sprint-worker
+
+# Start services
+sudo systemctl start ocr-sprint-api
+sudo systemctl start ocr-sprint-worker
+
+# Check status
+sudo systemctl status ocr-sprint-api
+sudo systemctl status ocr-sprint-worker
+
+# View logs
+sudo journalctl -u ocr-sprint-api -f
+sudo journalctl -u ocr-sprint-worker -f
+```
+
+## Langkah 9: Setup Nginx Reverse Proxy
+
+### Install Nginx
+
+```bash
+sudo apt install -y nginx certbot python3-certbot-nginx
+```
+
+### Konfigurasi Nginx
+
+```bash
+sudo nano /etc/nginx/sites-available/ocr-sprint
+```
+
+**Content `/etc/nginx/sites-available/ocr-sprint`:**
+
+```nginx
+# Upstream untuk load balancing (jika scale horizontal)
+upstream ocr_api {
+    server 127.0.0.1:8000;
+    keepalive 32;
+}
+
+# Rate limiting
+limit_req_zone $binary_remote_addr zone=api_limit:10m rate=10r/s;
+
+server {
+    listen 80;
+    server_name ocr.yourdomain.com;  # Ganti dengan domain Anda
+
+    # Max upload size (sesuaikan dengan BLOB_MAX_UPLOAD_MB)
+    client_max_body_size 30M;
+    client_body_buffer_size 128k;
+
+    # Timeouts untuk dokumen besar
+    proxy_connect_timeout 300s;
+    proxy_send_timeout 300s;
+    proxy_read_timeout 300s;
+    send_timeout 300s;
+
+    # Logging
+    access_log /var/log/nginx/ocr-sprint-access.log;
+    error_log /var/log/nginx/ocr-sprint-error.log;
+
+    # API endpoints
+    location /api/ {
+        # Rate limiting
+        limit_req zone=api_limit burst=20 nodelay;
+
+        proxy_pass http://ocr_api;
+        proxy_http_version 1.1;
+        
+        # Headers
+        proxy_set_header Host $host;
+        proxy_set_header X-Real-IP $remote_addr;
+        proxy_set_header X-Forwarded-For $proxy_add_x_forwarded_for;
+        proxy_set_header X-Forwarded-Proto $scheme;
+        proxy_set_header Connection "";
+        
+        # Disable buffering untuk streaming responses
+        proxy_buffering off;
+    }
+
+    # Health check endpoint (no rate limit)
+    location /api/v1/health {
+        proxy_pass http://ocr_api;
+        proxy_http_version 1.1;
+        proxy_set_header Host $host;
+        access_log off;
+    }
+
+    # Metrics endpoint (restrict access)
+    location /metrics {
+        # Allow only from internal network
+        allow 10.0.0.0/8;
+        allow 172.16.0.0/12;
+        allow 192.168.0.0/16;
+        allow 127.0.0.1;
+        deny all;
+
+        proxy_pass http://ocr_api;
+        proxy_http_version 1.1;
+        proxy_set_header Host $host;
+    }
+
+    # Docs (opsional, bisa di-disable di production)
+    location /docs {
+        proxy_pass http://ocr_api;
+        proxy_http_version 1.1;
+        proxy_set_header Host $host;
+    }
+
+    location /redoc {
+        proxy_pass http://ocr_api;
+        proxy_http_version 1.1;
+        proxy_set_header Host $host;
+    }
+}
+```
+
+### Enable Site
+
+```bash
+# Test konfigurasi
+sudo nginx -t
+
+# Enable site
+sudo ln -s /etc/nginx/sites-available/ocr-sprint /etc/nginx/sites-enabled/
+
+# Remove default site (opsional)
+sudo rm /etc/nginx/sites-enabled/default
+
+# Reload Nginx
+sudo systemctl reload nginx
+```
+
+### Setup SSL dengan Let's Encrypt
+
+```bash
+# Install certbot
+sudo apt install -y certbot python3-certbot-nginx
+
+# Obtain certificate (ganti dengan domain Anda)
+sudo certbot --nginx -d ocr.yourdomain.com
+
+# Test auto-renewal
+sudo certbot renew --dry-run
+```
+
+Certbot akan otomatis mengupdate konfigurasi Nginx untuk HTTPS.
+
+## Langkah 10: Setup Firewall
+
+```bash
+# Install UFW (jika belum ada)
+sudo apt install -y ufw
+
+# Allow SSH (PENTING! Jangan sampai terkunci)
+sudo ufw allow 22/tcp
+
+# Allow HTTP dan HTTPS
+sudo ufw allow 80/tcp
+sudo ufw allow 443/tcp
+
+# Enable firewall
+sudo ufw enable
+
+# Check status
+sudo ufw status
+```
+
+## Langkah 11: Verifikasi Deployment
+
+### Test dari Server
+
+```bash
+# Health check
+curl http://localhost:8000/api/v1/health
+
+# Test dengan API key
+curl -X POST http://localhost:8000/api/v1/documents?sync=true \
+  -H "X-API-Key: your-api-key-here" \
+  -F "file=@/path/to/test.pdf"
+```
+
+### Test dari Client
+
+```bash
+# Health check via domain
+curl https://ocr.yourdomain.com/api/v1/health
+
+# Upload dokumen
+curl -X POST https://ocr.yourdomain.com/api/v1/documents \
+  -H "X-API-Key: your-api-key-here" \
+  -F "file=@document.pdf"
+```
+
+## Monitoring dan Maintenance
+
+### View Logs
+
+```bash
+# API logs
+sudo journalctl -u ocr-sprint-api -f
+
+# Worker logs
+sudo journalctl -u ocr-sprint-worker -f
+
+# Nginx logs
+sudo tail -f /var/log/nginx/ocr-sprint-access.log
+sudo tail -f /var/log/nginx/ocr-sprint-error.log
+
+# PostgreSQL logs
+sudo tail -f /var/log/postgresql/postgresql-14-main.log
+```
+
+### Service Management
+
+```bash
+# Restart services
+sudo systemctl restart ocr-sprint-api
+sudo systemctl restart ocr-sprint-worker
+
+# Stop services
+sudo systemctl stop ocr-sprint-api
+sudo systemctl stop ocr-sprint-worker
+
+# Check status
+sudo systemctl status ocr-sprint-api
+sudo systemctl status ocr-sprint-worker
+```
+
+### Database Backup
+
+```bash
+# Create backup script
+sudo nano /opt/ocr-sprint-service/backup.sh
+```
+
+**Content `backup.sh`:**
+
+```bash
+#!/bin/bash
+BACKUP_DIR="/opt/ocr-sprint-service/backups"
+DATE=$(date +%Y%m%d_%H%M%S)
+
+mkdir -p $BACKUP_DIR
+
+# Backup database
+pg_dump -U ocr -h localhost ocr_sprint | gzip > $BACKUP_DIR/db_$DATE.sql.gz
+
+# Backup blobs (opsional, bisa besar)
+# tar -czf $BACKUP_DIR/blobs_$DATE.tar.gz /opt/ocr-sprint-service/storage/blobs
+
+# Keep only last 7 days
+find $BACKUP_DIR -name "db_*.sql.gz" -mtime +7 -delete
+
+echo "Backup completed: $DATE"
+```
+
+```bash
+# Make executable
+chmod +x /opt/ocr-sprint-service/backup.sh
+
+# Setup cron job (daily at 2 AM)
+sudo crontab -e
+
+# Add line:
+0 2 * * * /opt/ocr-sprint-service/backup.sh >> /var/log/ocr-backup.log 2>&1
+```
+
+### Log Rotation
+
+```bash
+sudo nano /etc/logrotate.d/ocr-sprint
+```
+
+**Content:**
+
+```
+/var/log/nginx/ocr-sprint-*.log {
+    daily
+    rotate 14
+    compress
+    delaycompress
+    notifempty
+    create 0640 www-data adm
+    sharedscripts
+    postrotate
+        [ -f /var/run/nginx.pid ] && kill -USR1 `cat /var/run/nginx.pid`
+    endscript
+}
+```
+
+## Update Application
+
+```bash
+# Switch ke user ocr
+sudo su - ocr
+cd /opt/ocr-sprint-service
+
+# Pull latest code
+git pull
+
+# Activate venv
+source .venv/bin/activate
+
+# Update dependencies
+pip install -e ".[ocr]"
+
+# Run migrations
+alembic upgrade head
+
+# Exit user ocr
+exit
+
+# Restart services
+sudo systemctl restart ocr-sprint-api
+sudo systemctl restart ocr-sprint-worker
+
+# Check logs
+sudo journalctl -u ocr-sprint-api -n 50
+```
+
+## Performance Tuning
+
+### Increase Worker Concurrency
+
+```bash
+# Edit worker service
+sudo nano /etc/systemd/system/ocr-sprint-worker.service
+
+# Ubah --concurrency sesuai CPU cores
+# Untuk 8 cores: --concurrency=4
+# Untuk 16 cores: --concurrency=8
+
+# Reload dan restart
+sudo systemctl daemon-reload
+sudo systemctl restart ocr-sprint-worker
+```
+
+### PostgreSQL Tuning
+
+```bash
+sudo nano /etc/postgresql/14/main/postgresql.conf
+```
+
+**Recommended settings untuk 16GB RAM:**
+
+```
+shared_buffers = 4GB
+effective_cache_size = 12GB
+maintenance_work_mem = 1GB
+checkpoint_completion_target = 0.9
+wal_buffers = 16MB
+default_statistics_target = 100
+random_page_cost = 1.1
+effective_io_concurrency = 200
+work_mem = 10MB
+min_wal_size = 1GB
+max_wal_size = 4GB
+max_worker_processes = 4
+max_parallel_workers_per_gather = 2
+max_parallel_workers = 4
+```
+
+```bash
+sudo systemctl restart postgresql
+```
+
+### Redis Tuning
+
+```bash
+sudo nano /etc/redis/redis.conf
+```
+
+**Recommended settings:**
+
+```
+maxmemory 2gb
+maxmemory-policy allkeys-lru
+save ""  # Disable RDB snapshots untuk performance
+```
+
+```bash
+sudo systemctl restart redis
+```
+
+## Troubleshooting
+
+### Service tidak start
+
+```bash
+# Check logs
+sudo journalctl -u ocr-sprint-api -n 100 --no-pager
+sudo journalctl -u ocr-sprint-worker -n 100 --no-pager
+
+# Check permissions
+ls -la /opt/ocr-sprint-service
+ls -la /opt/ocr-sprint-service/storage
+
+# Test manual run
+sudo su - ocr
+cd /opt/ocr-sprint-service
+source .venv/bin/activate
+uvicorn ocr_sprint.main:app --host 0.0.0.0 --port 8000
+```
+
+### Database connection error
+
+```bash
+# Test connection
+sudo -u ocr psql -h localhost -U ocr -d ocr_sprint
+
+# Check PostgreSQL status
+sudo systemctl status postgresql
+
+# Check pg_hba.conf
+sudo cat /etc/postgresql/14/main/pg_hba.conf | grep ocr
+```
+
+### Redis connection error
+
+```bash
+# Test Redis
+redis-cli ping
+
+# Check Redis status
+sudo systemctl status redis
+
+# Check Redis logs
+sudo journalctl -u redis -n 50
+```
+
+### PaddleOCR model download gagal
+
+```bash
+# Download manual
+sudo su - ocr
+cd /opt/ocr-sprint-service
+source .venv/bin/activate
+
+python << EOF
+from paddleocr import PaddleOCR
+ocr = PaddleOCR(use_angle_cls=True, lang='latin')
+print("Models downloaded successfully")
+EOF
+```
+
+### Out of memory
+
+```bash
+# Check memory usage
+free -h
+htop
+
+# Reduce worker concurrency
+sudo nano /etc/systemd/system/ocr-sprint-worker.service
+# Ubah --concurrency=1
+
+# Add swap (jika perlu)
+sudo fallocate -l 4G /swapfile
+sudo chmod 600 /swapfile
+sudo mkswap /swapfile
+sudo swapon /swapfile
+echo '/swapfile none swap sw 0 0' | sudo tee -a /etc/fstab
+```
+
+## Security Checklist
+
+- [ ] API keys diganti dengan nilai random yang kuat
+- [ ] Database password diganti dari default
+- [ ] Firewall enabled (UFW) - hanya port 22, 80, 443 terbuka
+- [ ] SSL/TLS enabled via Let's Encrypt
+- [ ] `/metrics` endpoint restricted ke internal network
+- [ ] Nginx rate limiting configured
+- [ ] PostgreSQL hanya listen di localhost
+- [ ] Redis hanya listen di localhost
+- [ ] Regular backup configured (cron job)
+- [ ] Log rotation configured
+- [ ] OS security updates enabled (`unattended-upgrades`)
+- [ ] Fail2ban installed untuk SSH protection
+
+## Monitoring dengan Prometheus (Opsional)
+
+### Install Prometheus
+
+```bash
+# Download Prometheus
+cd /tmp
+wget https://github.com/prometheus/prometheus/releases/download/v2.45.0/prometheus-2.45.0.linux-amd64.tar.gz
+tar xvfz prometheus-*.tar.gz
+sudo mv prometheus-2.45.0.linux-amd64 /opt/prometheus
+
+# Create user
+sudo useradd --no-create-home --shell /bin/false prometheus
+
+# Create directories
+sudo mkdir /etc/prometheus /var/lib/prometheus
+sudo chown prometheus:prometheus /var/lib/prometheus
+```
+
+### Configure Prometheus
+
+```bash
+sudo nano /etc/prometheus/prometheus.yml
+```
+
+**Content:**
+
+```yaml
+global:
+  scrape_interval: 15s
+
+scrape_configs:
+  - job_name: 'ocr-sprint'
+    static_configs:
+      - targets: ['localhost:8000']
+    metrics_path: '/metrics'
+```
+
+### Create Systemd Service
+
+```bash
+sudo nano /etc/systemd/system/prometheus.service
+```
+
+**Content:**
+
+```ini
+[Unit]
+Description=Prometheus
+After=network.target
+
+[Service]
+User=prometheus
+Group=prometheus
+Type=simple
+ExecStart=/opt/prometheus/prometheus \
+    --config.file=/etc/prometheus/prometheus.yml \
+    --storage.tsdb.path=/var/lib/prometheus/
+
+[Install]
+WantedBy=multi-user.target
+```
+
+```bash
+sudo systemctl daemon-reload
+sudo systemctl enable prometheus
+sudo systemctl start prometheus
+```
+
+Access Prometheus di `http://localhost:9090`
+
+## Support
+
+Untuk pertanyaan atau issues, hubungi tim development.

docs/DEPLOYMENT.md

@@ -0,0 +1,437 @@
+# Quickstart Deployment OCR Sprint Service
+
+Panduan deployment OCR Sprint Service ke server production untuk pemrosesan dokumen surat sprint Polri.
+
+## Prasyarat Server
+
+### Spesifikasi Minimum
+- **OS**: Linux (Ubuntu 20.04+ / Debian 11+ / RHEL 8+)
+- **CPU**: 4 cores (8 cores recommended untuk throughput tinggi)
+- **RAM**: 8 GB minimum (16 GB recommended)
+- **Storage**: 50 GB free space
+  - ~3 GB untuk model PaddleOCR
+  - ~1.5 GB untuk dependencies Python
+  - Sisanya untuk blob storage dokumen
+- **Network**: Port 8000 terbuka untuk API access
+
+### Software Requirements
+- Docker 24.0+ dan Docker Compose v2
+- Git
+- (Opsional) Nginx/Caddy untuk reverse proxy + SSL
+
+## Deployment dengan Docker Compose (Recommended)
+
+### 1. Clone Repository
+
+```bash
+# Login ke server sebagai user non-root dengan sudo access
+ssh user@your-server.com
+
+# Clone repository
+git clone https://github.com/Adriankf59/ocr-sprint-service.git
+cd ocr-sprint-service
+```
+
+### 2. Konfigurasi Environment
+
+```bash
+# Copy template environment
+cp .env.example .env
+
+# Edit konfigurasi production
+nano .env
+```
+
+**Konfigurasi penting untuk production:**
+
+```bash
+# ==== App ====
+APP_ENV=prod
+APP_LOG_LEVEL=INFO
+
+# ==== Storage ====
+STORAGE_LOCAL_DIR=/app/storage
+BLOB_STORAGE_DIR=/app/storage/blobs
+BLOB_MAX_UPLOAD_MB=25
+
+# ==== OCR ====
+OCR_LANG=latin
+OCR_USE_GPU=false              # set true jika server punya GPU NVIDIA
+OCR_MAX_IMAGE_SIDE=2200
+
+# ==== Preprocessing ====
+PREPROCESS_TARGET_DPI=300
+PREPROCESS_DENOISE=true
+PREPROCESS_DESKEW=true
+PREPROCESS_DETECT_DOCUMENT=true
+PREPROCESS_REMOVE_SHADOW=true
+
+# ==== Table Extraction ====
+TABLES_ENABLED=true
+
+# ==== Async Pipeline ====
+QUEUE_ENABLED=true
+REDIS_URL=redis://redis:6379/0
+CELERY_TASK_DEFAULT_QUEUE=ocr_sprint
+
+# ==== Database ====
+DATABASE_URL=postgresql+psycopg://ocr:ocr@postgres:5432/ocr_sprint
+DATABASE_ECHO=false
+
+# ==== Auth (WAJIB untuk production!) ====
+API_KEYS=your-secret-key-1,your-secret-key-2
+API_KEY_HEADER=X-API-Key
+```
+
+**Generate API keys yang aman:**
+
+```bash
+# Generate random API key
+openssl rand -hex 32
+```
+
+### 3. Build dan Start Services
+
+```bash
+# Build Docker images
+docker compose build
+
+# Start semua services (API, Worker, Redis, Postgres)
+docker compose up -d
+
+# Cek logs untuk memastikan semua berjalan
+docker compose logs -f api worker
+```
+
+**Services yang berjalan:**
+- `api`: FastAPI server di port 8000
+- `worker`: Celery worker untuk async processing
+- `redis`: Message broker untuk job queue
+- `postgres`: Database untuk job state
+
+### 4. Verifikasi Deployment
+
+```bash
+# Health check
+curl http://localhost:8000/api/v1/health
+
+# Expected response:
+# {"status":"ok","version":"0.1.0"}
+
+# Test OCR endpoint (sync mode untuk testing)
+curl -X POST http://localhost:8000/api/v1/documents?sync=true \
+  -H "X-API-Key: your-secret-key-1" \
+  -F "file=@samples/pdf/example.pdf" \
+  | jq
+```
+
+### 5. Setup Reverse Proxy (Nginx)
+
+**Install Nginx:**
+
+```bash
+sudo apt update
+sudo apt install nginx certbot python3-certbot-nginx
+```
+
+**Konfigurasi Nginx (`/etc/nginx/sites-available/ocr-sprint`):**
+
+```nginx
+upstream ocr_api {
+    server localhost:8000;
+}
+
+server {
+    listen 80;
+    server_name ocr.yourdomain.com;
+
+    client_max_body_size 30M;  # Sesuaikan dengan BLOB_MAX_UPLOAD_MB
+
+    location / {
+        proxy_pass http://ocr_api;
+        proxy_set_header Host $host;
+        proxy_set_header X-Real-IP $remote_addr;
+        proxy_set_header X-Forwarded-For $proxy_add_x_forwarded_for;
+        proxy_set_header X-Forwarded-Proto $scheme;
+        
+        # Timeout untuk dokumen besar
+        proxy_read_timeout 300s;
+        proxy_connect_timeout 75s;
+    }
+
+    location /metrics {
+        # Restrict metrics endpoint
+        allow 10.0.0.0/8;  # Internal network only
+        deny all;
+        proxy_pass http://ocr_api;
+    }
+}
+```
+
+**Enable site dan setup SSL:**
+
+```bash
+# Enable site
+sudo ln -s /etc/nginx/sites-available/ocr-sprint /etc/nginx/sites-enabled/
+sudo nginx -t
+sudo systemctl reload nginx
+
+# Setup SSL dengan Let's Encrypt
+sudo certbot --nginx -d ocr.yourdomain.com
+```
+
+## Deployment Manual (Tanpa Docker)
+
+### 1. Install System Dependencies
+
+```bash
+# Ubuntu/Debian
+sudo apt update
+sudo apt install -y \
+    python3.11 python3.11-venv python3-pip \
+    libgl1 libglib2.0-0 libsm6 libxext6 libxrender1 \
+    libgomp1 libmagic1 \
+    redis-server postgresql-14
+
+# Start services
+sudo systemctl enable --now redis-server postgresql
+```
+
+### 2. Setup Database
+
+```bash
+# Create database dan user
+sudo -u postgres psql << EOF
+CREATE USER ocr WITH PASSWORD 'your-secure-password';
+CREATE DATABASE ocr_sprint OWNER ocr;
+GRANT ALL PRIVILEGES ON DATABASE ocr_sprint TO ocr;
+EOF
+```
+
+### 3. Install Application
+
+```bash
+# Clone repository
+git clone https://github.com/Adriankf59/ocr-sprint-service.git
+cd ocr-sprint-service
+
+# Create virtual environment
+python3.11 -m venv .venv
+source .venv/bin/activate
+
+# Install dependencies
+pip install --upgrade pip
+pip install -e ".[ocr]"
+
+# Copy dan edit .env
+cp .env.example .env
+nano .env
+```
+
+**Update DATABASE_URL di .env:**
+
+```bash
+DATABASE_URL=postgresql+psycopg://ocr:your-secure-password@localhost:5432/ocr_sprint
+REDIS_URL=redis://localhost:6379/0
+QUEUE_ENABLED=true
+```
+
+### 4. Run Database Migrations
+
+```bash
+alembic upgrade head
+```
+
+### 5. Setup Systemd Services
+
+**API Service (`/etc/systemd/system/ocr-sprint-api.service`):**
+
+```ini
+[Unit]
+Description=OCR Sprint API
+After=network.target postgresql.service redis.service
+
+[Service]
+Type=simple
+User=ocr
+WorkingDirectory=/opt/ocr-sprint-service
+Environment="PATH=/opt/ocr-sprint-service/.venv/bin"
+ExecStart=/opt/ocr-sprint-service/.venv/bin/uvicorn ocr_sprint.main:app --host 0.0.0.0 --port 8000 --workers 4
+Restart=always
+RestartSec=10
+
+[Install]
+WantedBy=multi-user.target
+```
+
+**Worker Service (`/etc/systemd/system/ocr-sprint-worker.service`):**
+
+```ini
+[Unit]
+Description=OCR Sprint Celery Worker
+After=network.target postgresql.service redis.service
+
+[Service]
+Type=simple
+User=ocr
+WorkingDirectory=/opt/ocr-sprint-service
+Environment="PATH=/opt/ocr-sprint-service/.venv/bin"
+ExecStart=/opt/ocr-sprint-service/.venv/bin/celery -A ocr_sprint.worker.celery_app worker -l info --concurrency=2
+Restart=always
+RestartSec=10
+
+[Install]
+WantedBy=multi-user.target
+```
+
+**Enable dan start services:**
+
+```bash
+sudo systemctl daemon-reload
+sudo systemctl enable --now ocr-sprint-api ocr-sprint-worker
+sudo systemctl status ocr-sprint-api ocr-sprint-worker
+```
+
+## Monitoring dan Maintenance
+
+### Monitoring Logs
+
+```bash
+# Docker deployment
+docker compose logs -f api worker
+
+# Manual deployment
+sudo journalctl -u ocr-sprint-api -f
+sudo journalctl -u ocr-sprint-worker -f
+```
+
+### Prometheus Metrics
+
+Metrics tersedia di endpoint `/metrics`:
+
+```bash
+curl http://localhost:8000/metrics
+```
+
+**Key metrics:**
+- `ocr_documents_total`: Total dokumen diproses
+- `ocr_processing_duration_seconds`: Durasi processing
+- `ocr_confidence_score`: Distribusi confidence score
+- `celery_task_*`: Celery worker metrics
+
+### Backup Database
+
+```bash
+# Docker deployment
+docker compose exec postgres pg_dump -U ocr ocr_sprint > backup_$(date +%Y%m%d).sql
+
+# Manual deployment
+pg_dump -U ocr ocr_sprint > backup_$(date +%Y%m%d).sql
+```
+
+### Update Service
+
+```bash
+# Docker deployment
+cd ocr-sprint-service
+git pull
+docker compose build
+docker compose up -d
+
+# Manual deployment
+cd ocr-sprint-service
+git pull
+source .venv/bin/activate
+pip install -e ".[ocr]"
+alembic upgrade head
+sudo systemctl restart ocr-sprint-api ocr-sprint-worker
+```
+
+## Troubleshooting
+
+### Service tidak start
+
+```bash
+# Cek logs
+docker compose logs api worker
+
+# Cek health check
+curl http://localhost:8000/api/v1/health
+```
+
+### PaddleOCR model download gagal
+
+```bash
+# Download manual ke volume
+docker compose exec api python -c "from paddleocr import PaddleOCR; PaddleOCR(use_angle_cls=True, lang='latin')"
+```
+
+### Worker tidak memproses jobs
+
+```bash
+# Cek Redis connection
+docker compose exec worker redis-cli -h redis ping
+
+# Cek Celery worker status
+docker compose exec worker celery -A ocr_sprint.worker.celery_app inspect active
+```
+
+### Database migration error
+
+```bash
+# Cek current revision
+docker compose exec api alembic current
+
+# Force upgrade
+docker compose exec api alembic upgrade head
+```
+
+### Out of memory
+
+```bash
+# Kurangi worker concurrency di docker-compose.yml
+# Ubah: --concurrency=1 (default) atau tambahkan memory limit
+```
+
+## Security Checklist
+
+- [ ] API_KEYS diset dengan nilai random yang kuat
+- [ ] Firewall configured (hanya port 80/443 terbuka)
+- [ ] SSL/TLS enabled via Nginx + Let's Encrypt
+- [ ] Database password diganti dari default
+- [ ] `/metrics` endpoint restricted ke internal network
+- [ ] Regular backup database dan blob storage
+- [ ] Log rotation configured
+- [ ] OS security updates enabled
+
+## Performance Tuning
+
+### Untuk throughput tinggi:
+
+1. **Increase worker concurrency:**
+   ```yaml
+   # docker-compose.yml
+   command: ["celery", "-A", "ocr_sprint.worker.celery_app", "worker", "-l", "info", "--concurrency=4"]
+   ```
+
+2. **Scale workers horizontally:**
+   ```bash
+   docker compose up -d --scale worker=3
+   ```
+
+3. **Enable GPU (jika tersedia):**
+   ```bash
+   # .env
+   OCR_USE_GPU=true
+   ```
+
+4. **Tune Postgres:**
+   ```sql
+   -- Increase connection pool
+   ALTER SYSTEM SET max_connections = 200;
+   ALTER SYSTEM SET shared_buffers = '2GB';
+   ```
+
+## Support
+
+Untuk pertanyaan atau issues, hubungi tim development atau buat issue di repository.

docs/FRONTEND-INTEGRATION.md

@@ -0,0 +1,537 @@
+# Frontend Integration Guide
+
+Dokumen ini menjelaskan kontrak API yang perlu dipakai frontend untuk upload dokumen sprint, menampilkan hasil OCR, menjalankan review manual, dan approve hasil final.
+
+## Base URL
+
+Default local API:
+
+```text
+http://localhost:8000/api/v1
+```
+
+Untuk frontend, simpan URL di environment variable:
+
+```env
+VITE_OCR_API_BASE_URL=http://localhost:8000/api/v1
+```
+
+Jika `API_KEYS` di backend diisi, semua endpoint protected membutuhkan header:
+
+```http
+X-API-Key: <api-key>
+```
+
+Catatan: jangan expose API key production di frontend publik. Untuk deployment internal, gunakan reverse proxy atau session backend-for-frontend jika aksesnya tidak sepenuhnya trusted.
+
+## Health Check
+
+```http
+GET /health
+GET /health/ready
+```
+
+Contoh response `/health`:
+
+```json
+{
+  "status": "ok",
+  "version": "0.1.0"
+}
+```
+
+Contoh response `/health/ready`:
+
+```json
+{
+  "status": "ready",
+  "version": "0.1.0",
+  "models": {
+    "paddleocr": "ready",
+    "pp_structure": "disabled"
+  }
+}
+```
+
+Gunakan `/health/ready` untuk disable upload button sampai model OCR siap.
+
+## Upload Dokumen
+
+Endpoint:
+
+```http
+POST /documents
+POST /documents?sync=true
+```
+
+Body harus `multipart/form-data` dengan field `file`.
+
+Backend menerima PDF dan format image umum. Default max upload mengikuti backend config `BLOB_MAX_UPLOAD_MB`, saat ini 25 MB.
+
+### Recommended Flow
+
+Untuk frontend production, gunakan async flow:
+
+1. `POST /documents`
+2. Jika status HTTP `202`, ambil `job_id`
+3. Poll `GET /documents/{job_id}` setiap 1-3 detik
+4. Stop polling saat status `completed`, `needs_review`, atau `failed`
+
+Untuk local dev sederhana, `POST /documents?sync=true` boleh dipakai, tetapi request bisa lama karena OCR berjalan inline.
+
+### Upload Example
+
+```ts
+const API_BASE = import.meta.env.VITE_OCR_API_BASE_URL;
+const API_KEY = import.meta.env.VITE_OCR_API_KEY;
+
+async function uploadDocument(file: File) {
+  const form = new FormData();
+  form.append("file", file);
+
+  const res = await fetch(`${API_BASE}/documents`, {
+    method: "POST",
+    headers: API_KEY ? { "X-API-Key": API_KEY } : undefined,
+    body: form,
+  });
+
+  if (!res.ok) {
+    throw await readApiError(res);
+  }
+
+  return (await res.json()) as DocumentResponse;
+}
+```
+
+## Polling Job
+
+Endpoint:
+
+```http
+GET /documents/{job_id}
+```
+
+```ts
+const TERMINAL_STATUSES = new Set(["completed", "needs_review", "failed"]);
+
+async function getDocument(jobId: string) {
+  const res = await fetch(`${API_BASE}/documents/${jobId}`, {
+    headers: API_KEY ? { "X-API-Key": API_KEY } : undefined,
+  });
+
+  if (!res.ok) {
+    throw await readApiError(res);
+  }
+
+  return (await res.json()) as DocumentResponse;
+}
+
+async function pollDocument(jobId: string, onUpdate: (doc: DocumentResponse) => void) {
+  while (true) {
+    const doc = await getDocument(jobId);
+    onUpdate(doc);
+
+    if (TERMINAL_STATUSES.has(doc.status)) {
+      return doc;
+    }
+
+    await new Promise((resolve) => setTimeout(resolve, 2000));
+  }
+}
+```
+
+## Response Schema
+
+### DocumentResponse
+
+```ts
+type DocumentStatus =
+  | "pending"
+  | "processing"
+  | "completed"
+  | "needs_review"
+  | "failed";
+
+type DocumentResponse = {
+  job_id: string;
+  status: DocumentStatus;
+  confidence: number | null;
+  data: ExtractionResult | null;
+  review_flags: ReviewFlag[];
+  error: string | null;
+  approved: boolean;
+  reviewed_by: string | null;
+  reviewed_at: string | null;
+};
+```
+
+### ExtractionResult
+
+```ts
+type ExtractionResult = {
+  header: HeaderFields;
+  personel: PersonnelEntry[];
+  untuk: string[];
+  ttd: Signatory;
+  raw_text: string;
+  confidence: number;
+  review_flags: ReviewFlag[];
+};
+
+type HeaderFields = {
+  nomor_sprint: string | null;
+  tanggal: string | null; // YYYY-MM-DD
+  satuan_penerbit: string | null;
+  perihal: string | null;
+  dasar: string[];
+};
+
+type PersonnelEntry = {
+  no: number | null;
+  pangkat: string | null;
+  nrp: string | null;
+  nama: string | null;
+  jabatan_dinas: string | null;
+  jabatan_sprint: string | null;
+  keterangan: string | null;
+  confidence: number;
+};
+
+type Signatory = {
+  nama: string | null;
+  pangkat: string | null;
+  nrp: string | null;
+  jabatan: string | null;
+};
+```
+
+### Review Flags
+
+```ts
+type ReviewFlag =
+  | "low_ocr_confidence"
+  | "missing_field"
+  | "invalid_nrp"
+  | "unknown_pangkat"
+  | "personnel_count_mismatch"
+  | "date_parse_failed"
+  | "llm_fallback"
+  | "llm_unavailable"
+  | "personnel_text_fallback"
+  | "personnel_text_fallback_no_nrp"
+  | "incomplete_personnel_row";
+```
+
+Recommended UI labels:
+
+| Flag | Label |
+|---|---|
+| `low_ocr_confidence` | Confidence OCR rendah |
+| `missing_field` | Field wajib belum lengkap |
+| `invalid_nrp` | NRP tidak valid |
+| `unknown_pangkat` | Pangkat tidak dikenali |
+| `personnel_count_mismatch` | Jumlah personel perlu dicek |
+| `date_parse_failed` | Tanggal gagal dibaca |
+| `llm_fallback` | Sebagian field diisi fallback LLM |
+| `llm_unavailable` | LLM tidak tersedia |
+| `personnel_text_fallback` | Personel dibaca dari fallback teks |
+| `personnel_text_fallback_no_nrp` | Personel dibaca tanpa NRP |
+| `incomplete_personnel_row` | Baris personel belum lengkap |
+
+## Example Final Response
+
+```json
+{
+  "job_id": "e21e83ed-a42c-4672-baec-914e5c60cc5a",
+  "status": "needs_review",
+  "confidence": 0.82,
+  "data": {
+    "header": {
+      "nomor_sprint": "Sprin/123/IV/2026",
+      "tanggal": "2026-04-21",
+      "satuan_penerbit": "POLRES BANJAR",
+      "perihal": "Instruktur Ops Pekat I Lodaya 2026",
+      "dasar": []
+    },
+    "personel": [
+      {
+        "no": 1,
+        "pangkat": "IPDA",
+        "nrp": "12345678",
+        "nama": "BUDI SANTOSO",
+        "jabatan_dinas": "KANIT",
+        "jabatan_sprint": "INSTRUKTUR",
+        "keterangan": null,
+        "confidence": 0.91
+      }
+    ],
+    "untuk": ["Melaksanakan kegiatan sesuai surat perintah."],
+    "ttd": {
+      "nama": "AGUS",
+      "pangkat": "AKBP",
+      "nrp": "87654321",
+      "jabatan": "KAPOLRES"
+    },
+    "raw_text": "full OCR text...",
+    "confidence": 0.82,
+    "review_flags": ["low_ocr_confidence"]
+  },
+  "review_flags": ["low_ocr_confidence"],
+  "error": null,
+  "approved": false,
+  "reviewed_by": null,
+  "reviewed_at": null
+}
+```
+
+`raw_text` bisa panjang. Tampilkan di collapsible/debug panel, bukan di layar utama.
+
+## Review dan Koreksi HITL
+
+Frontend review screen sebaiknya mengizinkan editor untuk:
+
+- Header: nomor sprint, tanggal, satuan penerbit, perihal, dasar
+- Personel: pangkat, NRP, nama, jabatan dinas, jabatan sprint, keterangan
+- Untuk: daftar tugas
+- TTD: nama, pangkat, NRP, jabatan
+
+### Patch Corrections
+
+Endpoint:
+
+```http
+PATCH /documents/{job_id}
+```
+
+Body:
+
+```json
+{
+  "corrections": [
+    {
+      "path": "header.perihal",
+      "value": "Pelaksanaan Operasi Pekat I Lodaya 2026",
+      "reason": "OCR membaca perihal tidak lengkap"
+    },
+    {
+      "path": "personel[0].nama",
+      "value": "BUDI SANTOSO",
+      "reason": "Perbaikan nama"
+    }
+  ]
+}
+```
+
+Header opsional untuk audit trail:
+
+```http
+X-User-Id: reviewer-a
+```
+
+Path yang umum dipakai:
+
+```text
+header.nomor_sprint
+header.tanggal
+header.satuan_penerbit
+header.perihal
+header.dasar
+ttd.nama
+ttd.pangkat
+ttd.nrp
+ttd.jabatan
+personel[0].pangkat
+personel[0].nrp
+personel[0].nama
+personel[0].jabatan_dinas
+personel[0].jabatan_sprint
+personel[0].keterangan
+untuk
+```
+
+Semua correction dalam satu request bersifat atomic. Jika satu path invalid, seluruh batch ditolak dan tidak ada perubahan disimpan.
+
+### Patch Example
+
+```ts
+async function patchDocument(jobId: string, corrections: FieldCorrection[], userId?: string) {
+  const headers: Record<string, string> = { "Content-Type": "application/json" };
+  if (API_KEY) headers["X-API-Key"] = API_KEY;
+  if (userId) headers["X-User-Id"] = userId;
+
+  const res = await fetch(`${API_BASE}/documents/${jobId}`, {
+    method: "PATCH",
+    headers,
+    body: JSON.stringify({ corrections }),
+  });
+
+  if (!res.ok) {
+    throw await readApiError(res);
+  }
+
+  return (await res.json()) as DocumentResponse;
+}
+
+type FieldCorrection = {
+  path: string;
+  value: unknown;
+  reason?: string | null;
+};
+```
+
+## Correction History
+
+Endpoint:
+
+```http
+GET /documents/{job_id}/history
+```
+
+Response:
+
+```ts
+type CorrectionEventResponse = {
+  id: number;
+  job_id: string;
+  field_path: string;
+  old_value: unknown | null;
+  new_value: unknown | null;
+  corrected_by: string | null;
+  reason: string | null;
+  corrected_at: string;
+};
+```
+
+Gunakan endpoint ini untuk audit panel di halaman review.
+
+## Approve Final Result
+
+Endpoint:
+
+```http
+POST /documents/{job_id}/approve
+```
+
+Header opsional:
+
+```http
+X-User-Id: reviewer-a
+```
+
+Response:
+
+```json
+{
+  "job_id": "e21e83ed-a42c-4672-baec-914e5c60cc5a",
+  "approved": true,
+  "reviewed_by": "reviewer-a",
+  "reviewed_at": "2026-04-26T16:30:00"
+}
+```
+
+Setelah approved, `PATCH /documents/{job_id}` akan ditolak dengan `409`.
+
+## Error Handling
+
+Application errors:
+
+```json
+{
+  "error": "UnsupportedDocumentError",
+  "message": "Uploaded file is empty."
+}
+```
+
+FastAPI validation errors memakai shape standar:
+
+```json
+{
+  "detail": [
+    {
+      "type": "missing",
+      "loc": ["body", "file"],
+      "msg": "Field required"
+    }
+  ]
+}
+```
+
+Helper error:
+
+```ts
+async function readApiError(res: Response) {
+  let payload: unknown = null;
+  try {
+    payload = await res.json();
+  } catch {
+    payload = await res.text();
+  }
+
+  return {
+    status: res.status,
+    payload,
+  };
+}
+```
+
+Recommended UI handling:
+
+| HTTP Status | UI Handling |
+|---|---|
+| `400` | Tampilkan pesan validasi/upload |
+| `401` | Session/API key tidak valid |
+| `404` | Job tidak ditemukan |
+| `409` | Job belum selesai atau sudah approved |
+| `422` | Form correction tidak valid |
+| `500` | Tampilkan error umum dan minta operator cek log backend |
+
+## Ground Truth Admin
+
+Endpoint ini opsional untuk dashboard admin/training data.
+
+```http
+GET /ground-truth/stats?top_n=10
+GET /ground-truth/export?approved_only=true&has_corrections=true&limit=1000
+```
+
+`/ground-truth/export` mengembalikan `application/x-ndjson`, satu JSON per baris. Frontend biasanya cukup menyediakan tombol download, bukan parse seluruh stream di browser.
+
+## Recommended Screens
+
+1. Upload screen
+   - Dropzone file PDF/image
+   - Health readiness badge
+   - Upload progress
+   - Processing state setelah `job_id` diterima
+
+2. Result screen
+   - Status badge
+   - Confidence score
+   - Review flags
+   - Header summary
+   - Personnel table
+   - Untuk list
+   - TTD section
+   - Raw OCR collapsible
+
+3. Review screen
+   - Editable fields
+   - Dirty-state tracking
+   - Correction reason input
+   - Save corrections via `PATCH`
+   - History panel
+   - Approve button
+
+4. Admin screen
+   - Health/ready status
+   - Ground-truth stats
+   - Export approved samples
+
+## UX Rules
+
+- Jangan tunggu `POST /documents?sync=true` untuk production UI; gunakan async + polling.
+- Disable approve kalau status masih `pending` atau `processing`.
+- Tampilkan `needs_review` sebagai hasil yang berhasil diproses tetapi perlu validasi manusia.
+- Jangan render `raw_text` sebagai konten utama.
+- Pada `failed`, tampilkan `error` dari response jika ada.
+- Pada confidence rendah, arahkan user ke review fields yang punya flag terkait.

docs/OCR-RUNTIME-MODES.md

@@ -0,0 +1,49 @@
+# OCR Runtime Modes
+
+Backend OCR bisa dijalankan dalam mode CPU atau GPU lewat konfigurasi `OCR_USE_GPU`.
+
+## Cara Pakai
+
+Mode CPU:
+
+```powershell
+.\update.ps1 -OcrMode cpu
+```
+
+Mode GPU:
+
+```powershell
+.\update.ps1 -OcrMode gpu
+```
+
+Jika parameter tidak diberikan, `update.ps1` memakai nilai yang sudah ada di `.env`.
+
+```env
+OCR_USE_GPU=false
+```
+
+atau:
+
+```env
+OCR_USE_GPU=true
+```
+
+## Perilaku Script
+
+- `-OcrMode cpu` menyimpan `OCR_USE_GPU=false` ke `.env`.
+- `-OcrMode gpu` menyimpan `OCR_USE_GPU=true` ke `.env`.
+- Script tidak menghapus package Paddle/CUDA yang sudah terpasang.
+- Dalam mode GPU, script akan memasang `paddlepaddle-gpu` dan runtime cuDNN/cuBLAS jika belum ada.
+- Dalam mode CPU, script hanya memasang `paddlepaddle` CPU jika belum ada runtime Paddle sama sekali.
+
+## Catatan
+
+Mode CPU tidak membutuhkan CUDA, cuDNN, atau driver NVIDIA.
+
+Mode GPU membutuhkan NVIDIA driver dan runtime CUDA/cuDNN yang cocok. Pada Windows, backend juga menambahkan folder DLL NVIDIA dari `.venv` secara otomatis sebelum PaddleOCR diinisialisasi.
+
+`TABLES_ENABLED` adalah konfigurasi terpisah dari mode CPU/GPU. Jika PP-Structure belum stabil di environment lokal, biarkan:
+
+```env
+TABLES_ENABLED=false
+```

src/ocr_sprint/api/routes/documents.py

@@ -10,7 +10,10 @@ flow on top:
 * `POST /documents?sync=true` — runs the pipeline inline (the original
                                  Phase 1 behaviour). Useful for tests and
                                  small-volume single-tenant deploys without
-                                 a Celery worker.
+                                 a Celery worker. The heavy OCR work is
+                                 offloaded to a thread-pool executor so the
+                                 uvicorn event loop stays responsive during
+                                 processing (~30-120s on CPU).
 * `GET  /documents/{job_id}`  — returns the current job state. Async
                                  clients poll this until `status` is in a
                                  terminal state (completed / needs_review /
@@ -19,6 +22,9 @@ flow on top:
 
 from __future__ import annotations
 
+import asyncio
+from concurrent.futures import ThreadPoolExecutor
+from functools import partial
 from typing import Annotated
 from uuid import UUID, uuid4
 
@@ -60,6 +66,13 @@ from ocr_sprint.schemas.review import (
 from ocr_sprint.storage.blob import get_blob_storage
 from ocr_sprint.utils.logging import get_logger
 
+# Thread pool dedicated to blocking OCR work. Using a *separate* pool
+# (rather than the default loop executor) lets us cap the number of
+# concurrent heavy OCR jobs independently of other thread-pool users.
+# With 1 Celery worker + 1 sync slot we never exceed 2 parallel OCR
+# runs; keep the pool at 1 so RAM stays bounded on the 7.4 GB server.
+_OCR_EXECUTOR = ThreadPoolExecutor(max_workers=1, thread_name_prefix="ocr-inline")
+
 router = APIRouter(
     prefix="/documents",
     tags=["documents"],
@@ -86,9 +99,12 @@ def _row_to_response(row: object) -> DocumentResponse:
 
     assert isinstance(row, JobRow)
     status_enum = DocumentStatus(row.status)
-    result_obj: ExtractionResult | None = None
+    result_obj = None
     if row.result is not None:
         result_obj = ExtractionResult.model_validate(row.result)
+        # Auto-number personnel entries sequentially (1, 2, 3, ...)
+        for idx, entry in enumerate(result_obj.personel, start=1):
+            entry.no = idx
     return DocumentResponse(
         job_id=row.job_id,
         status=status_enum,
@@ -161,11 +177,13 @@ async def create_document(
 
 
 async def _run_inline(job_id: UUID, content: bytes) -> DocumentResponse:
-    """Synchronous pipeline execution.
+    """Run the OCR pipeline without blocking the uvicorn event loop.
 
-    Each state transition opens its own short session so the request-scoped
+    ``run_pipeline`` is CPU-bound and can take 30-120 s on a 2 vCPU server.
-    session's rollback-on-exception behaviour cannot wipe out the
+    Awaiting it directly on the async handler would freeze the entire event
-    ``mark_failed`` write or strand the blob on disk.
+    loop (and therefore block health-checks, metrics, and every other request)
+    for the full duration. We push the work onto a dedicated single-thread
+    executor so the loop stays free while the OCR runs in the background.
     """
     import time
 
@@ -173,8 +191,13 @@ async def _run_inline(job_id: UUID, content: bytes) -> DocumentResponse:
         JobRepository(s).mark_processing(job_id)
 
     started = time.perf_counter()
+    loop = asyncio.get_event_loop()
     try:
-        output = run_pipeline(content)
+        # run_pipeline is synchronous; wrap it so asyncio can await it.
+        output = await loop.run_in_executor(
+            _OCR_EXECUTOR,
+            partial(run_pipeline, content),
+        )
     except ValueError as exc:
         with session_scope() as s:
             JobRepository(s).mark_failed(job_id, error=str(exc))

src/ocr_sprint/api/routes/health.py

@@ -3,8 +3,12 @@
 from __future__ import annotations
 
 from fastapi import APIRouter
+from fastapi.responses import JSONResponse
 
 from ocr_sprint import __version__
+from ocr_sprint.config import get_settings
+from ocr_sprint.pipeline import ocr as _ocr
+from ocr_sprint.pipeline import table as _table
 
 router = APIRouter(tags=["health"])
 
@@ -13,3 +17,23 @@ router = APIRouter(tags=["health"])
 async def health() -> dict[str, str]:
     """Lightweight liveness check — does NOT touch the OCR engine."""
     return {"status": "ok", "version": __version__}
+
+
+@router.get("/health/ready")
+async def readiness() -> JSONResponse:
+    """Readiness check — returns 200 when OCR models are loaded, 503 if still warming up."""
+    settings = get_settings()
+    ocr_ready = _ocr._instance is not None
+    table_ready = (not settings.tables_enabled) or _table._instance is not None
+    ready = ocr_ready and table_ready
+    payload = {
+        "status": "ready" if ready else "warming_up",
+        "version": __version__,
+        "models": {
+            "paddleocr": "ready" if ocr_ready else "loading",
+            "pp_structure": (
+                "disabled" if not settings.tables_enabled else "ready" if table_ready else "loading"
+            ),
+        },
+    }
+    return JSONResponse(content=payload, status_code=200 if ready else 503)

src/ocr_sprint/config.py

@@ -24,6 +24,7 @@ class Settings(BaseSettings):
     app_host: str = "0.0.0.0"
     app_port: int = 8000
     app_log_level: str = "INFO"
+    root_path: str = ""  # For reverse proxy with path prefix (e.g., "/ocr")
 
     # Storage (Phase 1: local fs)
     storage_local_dir: Path = Path("./storage")

src/ocr_sprint/data/master_pangkat.py

@@ -22,7 +22,7 @@ PANGKAT_VARIANTS: dict[str, tuple[str, ...]] = {
     # Bintara
     "BRIPDA": ("BRIPDA",),
     "BRIPTU": ("BRIPTU",),
-    "BRIGADIR": ("BRIGADIR", "BRIG", "BRIG POL"),
+    "BRIGADIR": ("BRIGADIR", "BRIG", "BRIG POL", "BRIGPOL"),
     "BRIPKA": ("BRIPKA",),
     "AIPDA": ("AIPDA",),
     "AIPTU": ("AIPTU",),
@@ -33,12 +33,45 @@ PANGKAT_VARIANTS: dict[str, tuple[str, ...]] = {
     # Perwira Menengah
     "KOMPOL": ("KOMPOL",),
     "AKBP": ("AKBP",),
-    "KOMBES POL": ("KOMBES POL", "KOMBESPOL", "KBP"),
+    "KOMBES POL": ("KOMBES POL", "KOMBESPOL", "KBP", "KOMBES"),
     # Perwira Tinggi
     "BRIGJEN POL": ("BRIGJEN POL", "BRIGJENPOL", "BRIGJEN"),
     "IRJEN POL": ("IRJEN POL", "IRJENPOL", "IRJEN"),
     "KOMJEN POL": ("KOMJEN POL", "KOMJENPOL", "KOMJEN"),
     "JENDERAL POL": ("JENDERAL POL", "JENDERALPOL", "JENDERAL"),
+    # PNS Polri (Pegawai Negeri Sipil di lingkungan Polri). PNS appear
+    # routinely on sprint panitia / undangan templates alongside Polri
+    # personnel, so we treat them as valid ranks for extraction.
+    # Sources: PP 11/2017 jo PP 17/2020 (Manajemen PNS); golongan I-IV.
+    # Golongan I (Juru)
+    "JURU MUDA": ("JURU MUDA",),
+    "JURU MUDA TK I": ("JURU MUDA TK I", "JURU MUDA TK.I", "JURU MUDA TINGKAT I"),
+    "JURU": ("JURU",),
+    "JURU TK I": ("JURU TK I", "JURU TK.I", "JURU TINGKAT I"),
+    # Golongan II (Pengatur)
+    "PENGATUR MUDA": ("PENGATUR MUDA",),
+    "PENGATUR MUDA TK I": (
+        "PENGATUR MUDA TK I",
+        "PENGATUR MUDA TK.I",
+        "PENGATUR MUDA TINGKAT I",
+    ),
+    "PENGATUR": ("PENGATUR",),
+    "PENGATUR TK I": ("PENGATUR TK I", "PENGATUR TK.I", "PENGATUR TINGKAT I"),
+    # Golongan III (Penata)
+    "PENATA MUDA": ("PENATA MUDA",),
+    "PENATA MUDA TK I": (
+        "PENATA MUDA TK I",
+        "PENATA MUDA TK.I",
+        "PENATA MUDA TINGKAT I",
+    ),
+    "PENATA": ("PENATA",),
+    "PENATA TK I": ("PENATA TK I", "PENATA TK.I", "PENATA TINGKAT I"),
+    # Golongan IV (Pembina)
+    "PEMBINA": ("PEMBINA",),
+    "PEMBINA TK I": ("PEMBINA TK I", "PEMBINA TK.I", "PEMBINA TINGKAT I"),
+    "PEMBINA UTAMA MUDA": ("PEMBINA UTAMA MUDA",),
+    "PEMBINA UTAMA MADYA": ("PEMBINA UTAMA MADYA",),
+    "PEMBINA UTAMA": ("PEMBINA UTAMA",),
 }
 
 # Reverse lookup: any variant (uppercased) → canonical form.

src/ocr_sprint/ground_truth/service.py

@@ -171,7 +171,10 @@ def iter_ground_truth_samples(
             reviewed_at=job_row.reviewed_at,
             created_at=job_row.created_at,
             initial_result=initial,
-            final_result=copy.deepcopy(job_row.result) if job_row.result else None,
+            # Use an ``is None`` check to stay consistent with
+            # ``build_initial_result``; otherwise an empty-dict result
+            # would produce ``initial_result={}`` but ``final_result=None``.
+            final_result=(copy.deepcopy(job_row.result) if job_row.result is not None else None),
             corrections=[
                 GroundTruthCorrection(
                     field_path=c.field_path,

src/ocr_sprint/main.py

@@ -2,6 +2,10 @@
 
 from __future__ import annotations
 
+import threading
+from contextlib import asynccontextmanager
+from typing import AsyncIterator
+
 from fastapi import FastAPI
 
 from ocr_sprint import __version__
@@ -11,7 +15,10 @@ from ocr_sprint.api.routes import documents, ground_truth, health
 from ocr_sprint.config import get_settings
 from ocr_sprint.db import models as _models  # noqa: F401  (register ORM tables)
 from ocr_sprint.db.base import Base, get_engine
-from ocr_sprint.utils.logging import configure_logging
+from ocr_sprint.utils.logging import configure_logging, get_logger
+
+
+_startup_logger = get_logger(__name__)
 
 
 def _ensure_schema() -> None:
@@ -24,22 +31,74 @@ def _ensure_schema() -> None:
     Base.metadata.create_all(bind=get_engine())
 
 
+def _warmup_models_background() -> None:
+    """Load PaddleOCR and PP-Structure models in a background thread.
+
+    Running in a thread keeps the lifespan non-blocking so uvicorn can
+    start accepting health-check requests immediately while the heavy models
+    load (~5-15s on CPU). Requests that arrive before warmup completes will
+    wait on the existing _lock in each module rather than racing to load.
+    """
+    from ocr_sprint.config import get_settings as _gs
+    from ocr_sprint.pipeline import ocr as _ocr
+    from ocr_sprint.pipeline import table as _table
+
+    s = _gs()
+    try:
+        _ocr.warmup()
+    except Exception as exc:
+        _startup_logger.warning("paddleocr.warmup.failed", error=str(exc))
+
+    if s.tables_enabled:
+        try:
+            _table.warmup()
+        except Exception as exc:
+            _startup_logger.warning("pp_structure.warmup.failed", error=str(exc))
+
+
+@asynccontextmanager
+async def lifespan(app: FastAPI) -> AsyncIterator[None]:
+    """FastAPI lifespan: warm OCR models on startup in a background thread."""
+    _startup_logger.info("startup.warmup.begin")
+    t = threading.Thread(target=_warmup_models_background, name="ocr-warmup", daemon=True)
+    t.start()
+    yield
+    # Shutdown: nothing to clean up (models are process-global singletons).
+    _startup_logger.info("shutdown.complete")
+
+
 def create_app() -> FastAPI:
     """Application factory — keeps top-level state easy to test."""
     settings = get_settings()
     configure_logging(settings.app_log_level)
     _ensure_schema()
 
+    # Support for reverse proxy with path prefix (e.g., /ocr)
+    root_path = getattr(settings, "root_path", "")
+
     app = FastAPI(
+        lifespan=lifespan,
         title="OCR Sprint Service",
         version=__version__,
         description="OCR + structured extraction for Indonesian police 'surat sprint' documents.",
         docs_url="/docs",
         redoc_url="/redoc",
         openapi_url="/openapi.json",
+        root_path=root_path,
     )
 
     register_error_handlers(app)
+
+    # CORS — allow frontend dev servers and production origins
+    from fastapi.middleware.cors import CORSMiddleware
+    app.add_middleware(
+        CORSMiddleware,
+        allow_origins=["*"],
+        allow_credentials=True,
+        allow_methods=["*"],
+        allow_headers=["*"],
+    )
+
     app.add_middleware(MetricsMiddleware)
     app.include_router(health.router, prefix="/api/v1")
     app.include_router(documents.router, prefix="/api/v1")

src/ocr_sprint/pipeline/confidence.py

@@ -22,6 +22,14 @@ _FLAG_PENALTY: dict[ReviewFlag, float] = {
     ReviewFlag.UNKNOWN_PANGKAT: 0.05,
     ReviewFlag.PERSONNEL_COUNT_MISMATCH: 0.15,
     ReviewFlag.DATE_PARSE_FAILED: 0.10,
+    # Text-based personnel fallback is a recoverable degradation: rank/NRP
+    # were extracted via regex from raw OCR rather than from a parsed table
+    # grid. Worth flagging for review but not catastrophic.
+    ReviewFlag.PERSONNEL_TEXT_FALLBACK: 0.05,
+    # An incomplete personnel row (no pangkat AND no nrp) is a strong
+    # signal something went wrong. Penalise heavily so the document
+    # routes to needs_review even if the rest of the extraction is fine.
+    ReviewFlag.INCOMPLETE_PERSONNEL_ROW: 0.15,
 }
 
 OCR_WEIGHT = 0.6

src/ocr_sprint/pipeline/extract/personnel.py

@@ -64,6 +64,8 @@ _HEADER_SYNONYMS: dict[str, str] = {
     "jabatan dinas": "jabatan_dinas",
     "jabatan dalam dinas": "jabatan_dinas",
     "jbt dinas": "jabatan_dinas",
+    "struktural": "jabatan_dinas",
+    "jabatan struktural": "jabatan_dinas",
     # jabatan dalam sprint (role for this dispatch)
     "jabatan dalam sprint": "jabatan_sprint",
     "jabatan dalam sprin": "jabatan_sprint",
@@ -72,6 +74,8 @@ _HEADER_SYNONYMS: dict[str, str] = {
     "jabatan sprin": "jabatan_sprint",
     "tugas": "jabatan_sprint",
     "penugasan": "jabatan_sprint",
+    "dalam penugasan": "jabatan_sprint",
+    "jabatan dalam penugasan": "jabatan_sprint",
     # remarks
     "keterangan": "keterangan",
     "ket": "keterangan",

src/ocr_sprint/pipeline/extract/personnel_text.py

@@ -0,0 +1,797 @@
+"""Text-based fallback personnel extractor.
+
+PP-Structure (Phase 3) is the primary path for personnel rows because it
+preserves the table grid. But PP-Structure can fail in two ways on real
+sprint scans:
+
+1. The table is not detected at all (low-quality scan, watermark, atypical
+   layout) — `extract_personnel` returns an empty list.
+2. The table IS detected but the column mapping is too sparse, so each row
+   collapses to a single ``nama`` cell with all other fields ``None``. This
+   is what was observed on a real Polres Cimahi sprint where the OCR
+   produced 24 rows with only ``nama`` populated.
+
+This module provides a regex/heuristic fallback that operates directly on
+the flat OCR text. It is deliberately conservative: a row must have BOTH a
+recognizable Polri rank AND an 8-digit NRP to be emitted, so we never
+generate the kind of "name-only" rows that motivated the fallback in the
+first place.
+"""
+
+from __future__ import annotations
+
+import re
+
+from ocr_sprint.data.master_pangkat import (
+    PANGKAT_VARIANTS,
+    is_valid_pangkat,
+    normalize_pangkat,
+)
+from ocr_sprint.schemas.personnel import PersonnelEntry
+
+# Build a single alternation of all known rank tokens (longest first so multi-
+# word ranks like "KOMBES POL" win over the single-word "KOMBES").
+_RANK_TOKENS: tuple[str, ...] = tuple(
+    sorted(
+        {variant for variants in PANGKAT_VARIANTS.values() for variant in variants},
+        key=lambda v: -len(v),
+    )
+)
+_RANK_ALT = "|".join(re.escape(tok) for tok in _RANK_TOKENS)
+# A rank token followed (within a few characters) by an 8-digit NRP.
+# We allow common separators: '/', '-', '.', ',', ':' or whitespace.
+# The trailing ``\b`` plus proximity to the 8-digit NRP is the
+# specificity signal — we deliberately do *not* require a leading
+# ``\b`` because real Polri sprint OCR routinely mashes the rank into
+# the trailing characters of the previous cell (observed on Polres
+# Banjar: "...CPHR., CBA, CI" runs straight into "AKP" giving
+# "CIAKP 84011113"). Requiring a leading boundary loses that row
+# entirely. The longest-first alternation order ensures multi-token
+# ranks ("KOMBES POL") still win over short overlaps ("KBP").
+_RE_RANK_NRP_LINE = re.compile(
+    rf"(?P<rank>{_RANK_ALT})\b[\s/.\-,:]*?(?P<nrp>\d{{8}})\b",
+    re.IGNORECASE,
+)
+# A bare row number marker like "1." or "12)". OCR often puts it on its own
+# line in tabular layouts.
+_RE_ROW_NUMBER = re.compile(r"^\s*(\d{1,3})\s*[.)]\s*$")
+# Lines that should never be interpreted as a personnel name. These are
+# section headers, OCR garbage anchors, and column header tokens. We match
+# them with a *word-boundary* regex (built from this list) rather than a
+# bare ``startswith`` check, because short tokens like ``"NO"`` and
+# ``"KET"`` would otherwise reject perfectly valid Indonesian names
+# (e.g. ``"NOVA SARI"``, ``"NOOR HIDAYAT"``, ``"KETUT WARDANA"`` — the
+# latter being an extremely common Balinese birth-order name).
+_NAME_BLOCKLIST_TOKENS: tuple[str, ...] = (
+    "PADA TANGGAL",  # multi-word entries first so they win the alternation
+    "SURAT PERINTAH",
+    "DASAR",
+    "PERIHAL",
+    "PERTIMBANGAN",
+    "DIPERINTAHKAN",
+    "KEPADA",
+    "UNTUK",
+    "TEMBUSAN",
+    "DIKELUARKAN",
+    "SELESAI",
+    "DAFTAR",
+    "LAMPIRAN",
+    "NOMOR",
+    "TANGGAL",
+    "KEPOLISIAN",
+    "DAERAH",
+    "RESOR",
+    "SEKTOR",
+    "MABES",
+    "NRP",
+    "NIP",
+    "PANGKAT",
+    "JABATAN",
+    "NAMA",
+    "KETERANGAN",
+    "KET",
+    "NO",
+)
+_RE_NAME_BLOCKLIST = re.compile(
+    r"^(?:" + "|".join(re.escape(tok) for tok in _NAME_BLOCKLIST_TOKENS) + r")\b",
+    re.IGNORECASE,
+)
+# A name should look like a name: mostly letters, common punctuation, and
+# at least one alphabetic character. Pure-numeric or pure-symbol lines are
+# rejected.
+_RE_NAME_OK = re.compile(r"[A-Za-z]")
+
+
+def _is_plausible_name(line: str) -> bool:
+    """Return True iff ``line`` could plausibly be a personnel name."""
+    stripped = line.strip()
+    if not stripped or not _RE_NAME_OK.search(stripped):
+        return False
+    if _RE_NAME_BLOCKLIST.match(stripped):
+        return False
+    if _RE_ROW_NUMBER.match(stripped):
+        return False
+    if _RE_RANK_NRP_LINE.search(stripped):
+        return False
+    # Reject lines that are nothing but a row number with extra punctuation
+    # ("1 .", "2)") which the bare-number regex above might miss.
+    return not re.fullmatch(r"[\s\d.)(\-]+", stripped)
+
+
+def _following_jabatan(lines: list[str], idx: int) -> str | None:
+    """Collect 1-3 follow-up lines after the rank+NRP line as the jabatan.
+
+    Stops at the next rank+NRP line, the next bare row-number line, or any
+    blocked prefix (section header / column header).
+    """
+    parts: list[str] = []
+    for fwd in range(idx + 1, min(idx + 4, len(lines))):
+        candidate = lines[fwd].strip()
+        if not candidate:
+            if parts:
+                break
+            continue
+        if _RE_RANK_NRP_LINE.search(candidate):
+            break
+        if _RE_ROW_NUMBER.match(candidate):
+            break
+        if _RE_NAME_BLOCKLIST.match(candidate):
+            break
+        parts.append(candidate)
+    if not parts:
+        return None
+    joined = " ".join(parts)
+    return " ".join(joined.split()) or None
+
+
+def extract_personnel_from_text(raw_text: str) -> list[PersonnelEntry]:
+    """Best-effort personnel extraction from a flat OCR text stream.
+
+    Strategy:
+
+    **Pass 1** — same-line rank+NRP (original strategy):
+    1. Iterate every line. Skip lines that don't contain both a known rank
+       and an 8-digit NRP (those are the only signal we trust).
+    2. For each rank+NRP line, look back for the most recent plausible name
+       line, and forward 1-3 lines for jabatan content.
+    3. Emit a ``PersonnelEntry`` only when we have at least pangkat + nrp.
+
+    **Pass 2** — separate-line rank and NRP (for tabular sprint formats):
+    If pass 1 produces no results, scan for lines containing a standalone
+    rank token, then look up to 2 lines forward for a standalone NRP.
+    This handles sprint formats where OCR renders each column on its own
+    line (e.g. Polres Banjar layout).
+
+    **Pass 3** — rank-only (for sprint formats *without* an NRP column):
+    Some sprint templates (panitia, undangan, etc.) list only nama +
+    pangkat + jabatan, no NRP. If pass 1 and pass 2 both yield nothing,
+    fall back to a rank-only scan: every standalone rank line (or
+    two-line rank like "KOMBES" + "POL" produced by narrow-column OCR)
+    becomes a row, with name assembled from preceding lines and jabatan
+    from following lines. ``nrp`` stays ``None``. False-positive risk
+    is higher (stray rank tokens in body text), so this only fires when
+    nothing else matched.
+
+    The fallback is intentionally rate-limited: the first matching rank
+    token on a line wins (no greedy multi-match per line), and a name line
+    can only be consumed once (so a stray ranked text inside a paragraph
+    doesn't turn into multiple bogus entries).
+    """
+    lines = raw_text.splitlines()
+
+    # ── Pass 1: rank+NRP on the same line ────────────────────────────
+    rows = _extract_same_line(lines)
+    if rows:
+        return rows
+
+    # ── Pass 2: rank and NRP on separate lines ───────────────────────
+    rows = _extract_separate_lines(lines)
+    if rows:
+        return rows
+
+    # ── Pass 3: rank-only (no NRP column) ────────────────────────────
+    return _extract_rank_only(lines)
+
+
+# Regex for a line that is *only* a rank token (possibly with punctuation).
+_RE_RANK_ONLY = re.compile(
+    rf"^\s*(?P<rank>{_RANK_ALT})\s*[/.\-,:]*\s*$",
+    re.IGNORECASE,
+)
+# Regex for a line that contains a standalone 8-digit NRP.
+_RE_NRP_ONLY = re.compile(r"(?<!\d)(?P<nrp>\d{8})(?!\d)")
+
+
+# Strip a leading row number marker like "1 ", "1.", "12)" from a name
+# prefix taken from the same OCR line as a rank+NRP match. Unlike
+# _RE_ROW_NUMBER (which matches a *whole* line), this is a prefix strip
+# for embedded same-line cases like "1 CUCU JUHANA, A.K.S. KOMPOL ...".
+_RE_LEADING_ROW_NUMBER = re.compile(r"^\s*\d{1,3}\s*[.):]?\s+")
+
+
+def _extract_same_line(lines: list[str]) -> list[PersonnelEntry]:
+    """Pass 1: rank+NRP pairs found anywhere in the joined text.
+
+    Uses ``finditer`` over the full ``\\n``-joined OCR text rather than
+    ``re.search`` per line so that multiple rank+NRP pairs on the same
+    OCR line still produce separate rows. This is required for sprint
+    scans where Paddle merges several table rows into one OCR line
+    (observed on Polres Banjar where row 2's "...CBA.AKP 77020049 KASAT
+    RESKRIM" was being swallowed into row 1's jabatan because per-line
+    ``search`` only returns the first match).
+
+    For each match we resolve nama from text *before* the match (the
+    same-line prefix takes precedence; otherwise look back through the
+    preceding lines bounded by the previous match) and jabatan from text
+    *after* the match (same-line suffix plus up to ~3 follow-up lines,
+    bounded by the next match).
+    """
+    if not lines:
+        return []
+    full_text = "\n".join(lines)
+
+    line_starts: list[int] = []
+    pos = 0
+    for line in lines:
+        line_starts.append(pos)
+        pos += len(line) + 1  # +1 for the joining "\n"
+
+    def offset_to_line(offset: int) -> int:
+        lo, hi = 0, len(line_starts)
+        while lo < hi:
+            mid = (lo + hi) // 2
+            if line_starts[mid] <= offset:
+                lo = mid + 1
+            else:
+                hi = mid
+        return max(0, lo - 1)
+
+    matches = list(_RE_RANK_NRP_LINE.finditer(full_text))
+    rows: list[PersonnelEntry] = []
+    consumed_lines: set[int] = set()
+
+    for i, m in enumerate(matches):
+        pangkat = normalize_pangkat(m.group("rank"))
+        if not pangkat or not is_valid_pangkat(pangkat):
+            continue
+        nrp = m.group("nrp")
+        ml = offset_to_line(m.start())
+        prev_ml = (
+            offset_to_line(matches[i - 1].start()) if i > 0 else -1
+        )
+        next_ml = (
+            offset_to_line(matches[i + 1].start())
+            if i + 1 < len(matches)
+            else len(lines)
+        )
+
+        line_text = lines[ml]
+        line_off = line_starts[ml]
+
+        # Same-line prefix: text on this line *before* the rank token.
+        # If the previous match was on this same line, only consider the
+        # text after that previous match's NRP (otherwise we'd reuse the
+        # earlier row's tail as this row's name).
+        prefix_start_local = 0
+        if prev_ml == ml and i > 0:
+            prefix_start_local = max(0, matches[i - 1].end() - line_off)
+        prefix = line_text[prefix_start_local : m.start() - line_off]
+
+        # Same-line suffix: text on this line *after* the NRP, capped at
+        # the next match's start if it's on this same line.
+        suffix_end_local = len(line_text)
+        if next_ml == ml and i + 1 < len(matches):
+            suffix_end_local = matches[i + 1].start() - line_off
+        suffix = line_text[m.end() - line_off : suffix_end_local]
+
+        # ── Resolve nama ────────────────────────────────────────────
+        nama: str | None = None
+        prefix_clean = _RE_LEADING_ROW_NUMBER.sub("", prefix).strip()
+        if prefix_clean and _is_plausible_name(prefix_clean):
+            nama = prefix_clean
+        elif prev_ml < ml:
+            for back in range(ml - 1, prev_ml, -1):
+                if back in consumed_lines or back < 0:
+                    continue
+                candidate = lines[back].strip()
+                if _is_plausible_name(candidate):
+                    nama = candidate
+                    consumed_lines.add(back)
+                    break
+
+        # ── Resolve jabatan ─────────────────────────────────────────
+        jabatan_parts: list[str] = []
+        suffix_clean = suffix.strip()
+        if suffix_clean:
+            jabatan_parts.append(suffix_clean)
+        if next_ml > ml:
+            max_fwd = min(ml + 4, next_ml, len(lines))
+            for fwd in range(ml + 1, max_fwd):
+                candidate = lines[fwd].strip()
+                if not candidate:
+                    if jabatan_parts:
+                        break
+                    continue
+                if _RE_NAME_BLOCKLIST.match(candidate):
+                    break
+                if _RE_ROW_NUMBER.match(candidate):
+                    break
+                jabatan_parts.append(candidate)
+        jabatan = (
+            " ".join(" ".join(jabatan_parts).split())
+            if jabatan_parts
+            else None
+        )
+
+        rows.append(
+            PersonnelEntry(
+                no=None,
+                pangkat=pangkat,
+                nrp=nrp,
+                nama=nama,
+                jabatan_dinas=jabatan,
+                jabatan_sprint=None,
+                keterangan=None,
+            )
+        )
+    return rows
+
+
+def _extract_separate_lines(lines: list[str]) -> list[PersonnelEntry]:
+    """Pass 2: rank and NRP on separate nearby lines.
+
+    Handles tabular sprint formats where OCR outputs each column as its
+    own line, e.g.:
+        1
+        CUCU JUHANA, A.K.S.
+        KOMPOL
+        70100418
+        KABAGOPS
+    """
+    consumed_names: set[int] = set()
+    consumed_nrps: set[int] = set()
+    rows: list[PersonnelEntry] = []
+
+    for idx, raw_line in enumerate(lines):
+        line = raw_line.strip()
+        rank_match = _RE_RANK_ONLY.match(line)
+        if not rank_match:
+            # Also try: line starts with a rank token (may have trailing text)
+            for tok in _RANK_TOKENS:
+                if line.upper().startswith(tok) and len(line) - len(tok) < 5:
+                    rank_match = re.match(
+                        rf"^\s*(?P<rank>{re.escape(tok)})\s*[/.\-,:]*",
+                        line,
+                        re.IGNORECASE,
+                    )
+                    if rank_match:
+                        break
+        if not rank_match:
+            continue
+
+        pangkat = normalize_pangkat(rank_match.group("rank"))
+        if not pangkat or not is_valid_pangkat(pangkat):
+            continue
+
+        # Look forward up to 2 lines for NRP
+        nrp: str | None = None
+        nrp_idx: int | None = None
+        for fwd in range(idx + 1, min(idx + 3, len(lines))):
+            if fwd in consumed_nrps:
+                continue
+            nrp_match = _RE_NRP_ONLY.search(lines[fwd].strip())
+            if nrp_match:
+                nrp = nrp_match.group("nrp")
+                nrp_idx = fwd
+                break
+
+        if not nrp:
+            continue
+        assert nrp_idx is not None
+        consumed_nrps.add(nrp_idx)
+
+        # Look back for name
+        nama: str | None = None
+        for back in range(idx - 1, max(idx - 6, -1), -1):
+            if back in consumed_names:
+                continue
+            candidate = lines[back].strip()
+            if _is_plausible_name(candidate):
+                nama = candidate
+                consumed_names.add(back)
+                break
+
+        # Look forward after NRP for jabatan
+        jabatan = _following_jabatan(lines, nrp_idx)
+        rows.append(
+            PersonnelEntry(
+                no=None,
+                pangkat=pangkat,
+                nrp=nrp,
+                nama=nama,
+                jabatan_dinas=jabatan,
+                jabatan_sprint=None,
+                keterangan=None,
+            )
+        )
+    return rows
+
+
+# Bare row-number markers used by sprint formats without NRP (the dot
+# is often missing in narrow-column OCR, e.g. just "1" on its own line).
+_RE_BARE_ROW_NUMBER = re.compile(r"^\s*\d{1,3}\s*[.):]?\s*$")
+
+
+def _try_match_rank_at(lines: list[str], idx: int) -> tuple[str, int] | None:
+    """Try to match a standalone rank starting at ``lines[idx]``.
+
+    Returns ``(rank_text, lines_consumed)`` on success. Handles narrow-
+    column OCR that splits a multi-token rank across two lines (e.g.
+    ``"KOMBES"`` + ``"POL"`` or ``"PENATA"`` + ``"TK I"``).
+
+    The two-line concatenation is tried *first* so that more-specific
+    multi-token ranks ("PENATA TK I") win over their less-specific
+    single-line prefix ("PENATA"). Without this preference, "TK I"
+    would leak into the jabatan column.
+    """
+    if idx >= len(lines):
+        return None
+    line = lines[idx].strip()
+    if idx + 1 < len(lines):
+        combined = (line + " " + lines[idx + 1].strip()).strip()
+        m2 = _RE_RANK_ONLY.match(combined)
+        if m2:
+            return m2.group("rank"), 2
+    m = _RE_RANK_ONLY.match(line)
+    if m:
+        return m.group("rank"), 1
+    return None
+
+
+def _extract_rank_only(lines: list[str]) -> list[PersonnelEntry]:
+    """Pass 3: rank-only fallback for sprint formats without an NRP column.
+
+    Each standalone rank line (single line or two-line concatenation) is
+    treated as the pivot of a personnel row. ``nama`` is assembled from
+    the preceding contiguous plausible-name lines (typical OCR splits a
+    long name across 2-3 short lines because of narrow columns); jabatan
+    is collected from following lines until the next rank or row marker.
+
+    ``nrp`` is always ``None`` for rows produced by this pass.
+    """
+    rows: list[PersonnelEntry] = []
+    consumed_lines: set[int] = set()
+    i = 0
+    while i < len(lines):
+        match = _try_match_rank_at(lines, i)
+        if not match:
+            i += 1
+            continue
+        rank_text, rank_span = match
+        pangkat = normalize_pangkat(rank_text)
+        if not pangkat or not is_valid_pangkat(pangkat):
+            i += 1
+            continue
+
+        # ── Look back for name lines (assemble up to 4 contiguous lines) ──
+        name_lines: list[str] = []
+        for back in range(i - 1, max(i - 6, -1), -1):
+            if back in consumed_lines:
+                break
+            candidate = lines[back].strip()
+            if not candidate:
+                if name_lines:
+                    break
+                continue
+            if _RE_BARE_ROW_NUMBER.match(candidate):
+                break
+            if _RE_NAME_BLOCKLIST.match(candidate):
+                break
+            if _try_match_rank_at(lines, back) is not None:
+                break
+            if not _is_plausible_name(candidate):
+                break
+            name_lines.insert(0, candidate)
+            consumed_lines.add(back)
+        nama = " ".join(" ".join(name_lines).split()) if name_lines else None
+
+        # ── Look forward for jabatan (stop at next rank / row marker) ─────
+        jabatan_parts: list[str] = []
+        fwd = i + rank_span
+        steps = 0
+        while fwd < len(lines) and steps < 8:
+            candidate = lines[fwd].strip()
+            if not candidate:
+                if jabatan_parts:
+                    break
+                fwd += 1
+                steps += 1
+                continue
+            if _RE_BARE_ROW_NUMBER.match(candidate):
+                break
+            if _try_match_rank_at(lines, fwd) is not None:
+                break
+            if _RE_NAME_BLOCKLIST.match(candidate):
+                break
+            jabatan_parts.append(candidate)
+            fwd += 1
+            steps += 1
+        jabatan = " ".join(" ".join(jabatan_parts).split()) if jabatan_parts else None
+
+        rows.append(
+            PersonnelEntry(
+                no=None,
+                pangkat=pangkat,
+                nrp=None,
+                nama=nama,
+                jabatan_dinas=jabatan,
+                jabatan_sprint=None,
+                keterangan=None,
+            )
+        )
+        i += rank_span
+    return rows
+
+
+# ── Column-aware Pass 3 (uses OCR bounding boxes) ───────────────────────
+
+
+def _box_x_left(box: tuple[tuple[float, float], ...]) -> float:
+    return min(p[0] for p in box)
+
+
+def _box_x_right(box: tuple[tuple[float, float], ...]) -> float:
+    return max(p[0] for p in box)
+
+
+def _box_x_center(box: tuple[tuple[float, float], ...]) -> float:
+    return (_box_x_left(box) + _box_x_right(box)) / 2
+
+
+def _box_y_top(box: tuple[tuple[float, float], ...]) -> float:
+    return min(p[1] for p in box)
+
+
+def _box_y_bottom(box: tuple[tuple[float, float], ...]) -> float:
+    return max(p[1] for p in box)
+
+
+def _box_y_center(box: tuple[tuple[float, float], ...]) -> float:
+    return (_box_y_top(box) + _box_y_bottom(box)) / 2
+
+
+def _box_height(box: tuple[tuple[float, float], ...]) -> float:
+    return _box_y_bottom(box) - _box_y_top(box)
+
+
+def extract_personnel_from_ocr_lines(ocr_lines: list) -> list[PersonnelEntry]:
+    """Column-aware Pass 3 for sprint formats without an NRP column.
+
+    Each ``ocr_line`` must expose ``text`` (str) and ``box`` (a tuple of
+    4 ``(x, y)`` corner points). We use the geometry to:
+
+    1. Detect rank lines (single-line or vertically-stacked two-line).
+    2. Estimate the PANGKAT column X-center from those rank lines.
+    3. For each rank, gather **only** lines in the NAMA column (X left
+       of PANGKAT) within the row's Y span as the name fragments, and
+       **only** lines in the JABATAN column (X right of PANGKAT) for
+       jabatan. This prevents column-bleed that flat-text Pass 3
+       suffers from on dense tables.
+
+    Returns ``[]`` if no rank lines are detected (caller can fall back
+    to the text-only Pass 3).
+    """
+    if not ocr_lines:
+        return []
+
+    # Sort by (y_top, x_left) for vertical-stacking rank detection.
+    indexed = sorted(
+        range(len(ocr_lines)),
+        key=lambda i: (_box_y_top(ocr_lines[i].box), _box_x_left(ocr_lines[i].box)),
+    )
+
+    # Pass 1: find rank anchors.
+    # An anchor is one or two stacked OCR lines whose combined text matches
+    # _RE_RANK_ONLY and normalises to a known pangkat. Two-line stacks must
+    # X-overlap so we don't accidentally merge cells from different columns.
+    used: set[int] = set()
+    anchors: list[dict] = []
+    for pos, idx in enumerate(indexed):
+        if idx in used:
+            continue
+        ln = ocr_lines[idx]
+        text = ln.text.strip()
+
+        rank_text: str | None = None
+        member_idxs: list[int] = [idx]
+
+        # Try two-line stack first (so PENATA TK I beats PENATA).
+        for j_pos in range(pos + 1, min(pos + 5, len(indexed))):
+            j_idx = indexed[j_pos]
+            if j_idx in used:
+                continue
+            other = ocr_lines[j_idx]
+            x_overlap = (
+                min(_box_x_right(ln.box), _box_x_right(other.box))
+                - max(_box_x_left(ln.box), _box_x_left(other.box))
+            )
+            if x_overlap <= 0:
+                continue
+            y_gap = _box_y_top(other.box) - _box_y_bottom(ln.box)
+            if y_gap > _box_height(ln.box) * 1.5:
+                break
+            combined = (text + " " + other.text.strip()).strip()
+            m2 = _RE_RANK_ONLY.match(combined)
+            if m2:
+                rank_text = m2.group("rank")
+                member_idxs.append(j_idx)
+                break
+
+        if rank_text is None:
+            m1 = _RE_RANK_ONLY.match(text)
+            if m1:
+                rank_text = m1.group("rank")
+
+        if rank_text is None:
+            continue
+        pangkat = normalize_pangkat(rank_text)
+        if not pangkat or not is_valid_pangkat(pangkat):
+            continue
+
+        anchors.append(
+            {
+                "member_idxs": member_idxs,
+                "pangkat": pangkat,
+                "x_center": _box_x_center(ln.box),
+                "y_top": min(_box_y_top(ocr_lines[m].box) for m in member_idxs),
+                "y_bottom": max(_box_y_bottom(ocr_lines[m].box) for m in member_idxs),
+            }
+        )
+        used.update(member_idxs)
+
+    if not anchors:
+        return []
+
+    # Sort anchors by Y so we can compute row spans.
+    anchors.sort(key=lambda a: a["y_top"])
+
+    # Estimate PANGKAT column X-center as the median of rank anchor X-centers.
+    xs_sorted = sorted(a["x_center"] for a in anchors)
+    pangkat_x = xs_sorted[len(xs_sorted) // 2]
+
+    # X tolerance: half the median rank-line width. Lines with x_center
+    # within ±tolerance of pangkat_x are *in* the PANGKAT column and
+    # excluded from both NAMA and JABATAN buckets.
+    rank_widths = [
+        _box_x_right(ocr_lines[a["member_idxs"][0]].box)
+        - _box_x_left(ocr_lines[a["member_idxs"][0]].box)
+        for a in anchors
+    ]
+    rank_widths.sort()
+    median_rank_width = rank_widths[len(rank_widths) // 2] if rank_widths else 50.0
+    column_margin = max(median_rank_width * 0.5, 5.0)
+
+    # Try to split the JABATAN side into STRUKTURAL (jabatan_dinas) and
+    # DALAM SPRIN (jabatan_sprint) by clustering jabatan-side X-centers.
+    # This is a 2-cluster k-means-style split: collect all X-centers of
+    # lines to the right of PANGKAT, find the largest X-gap among them,
+    # and use that gap as the column boundary. KET is typically the
+    # right-most narrow column we let bleed into jabatan_sprint since
+    # it's commonly empty.
+    jabatan_xs: list[float] = []
+    for ln in ocr_lines:
+        x = _box_x_center(ln.box)
+        if x > pangkat_x + column_margin and ln.text.strip():
+            jabatan_xs.append(x)
+    jabatan_split_x: float | None = None
+    if len(jabatan_xs) >= 4:
+        jabatan_xs.sort()
+        max_gap = 0.0
+        max_gap_x: float | None = None
+        for k in range(1, len(jabatan_xs)):
+            gap = jabatan_xs[k] - jabatan_xs[k - 1]
+            if gap > max_gap:
+                max_gap = gap
+                max_gap_x = (jabatan_xs[k] + jabatan_xs[k - 1]) / 2
+        # Only use the split if the gap is meaningfully larger than a
+        # within-column gap (heuristic: > 1.5× median rank width).
+        if max_gap_x is not None and max_gap > median_rank_width * 1.5:
+            jabatan_split_x = max_gap_x
+
+    # Pre-compute each anchor's y_center for midpoint row dividers.
+    anchor_y_centers = [(a["y_top"] + a["y_bottom"]) / 2 for a in anchors]
+
+    rows: list[PersonnelEntry] = []
+    for i, anchor in enumerate(anchors):
+        # Row Y span: midpoint between this anchor and its neighbours.
+        # Using the midpoint (rather than the previous anchor's
+        # y_bottom) prevents row N's tail content (e.g. last name
+        # fragment "M.H.") from leaking into row N+1's nama bucket
+        # when rank lines don't extend to the full visual row height.
+        y_lo = (
+            (anchor_y_centers[i - 1] + anchor_y_centers[i]) / 2
+            if i > 0
+            else float("-inf")
+        )
+        y_hi = (
+            (anchor_y_centers[i] + anchor_y_centers[i + 1]) / 2
+            if i + 1 < len(anchors)
+            else float("inf")
+        )
+
+        nama_pieces: list[tuple[float, str]] = []
+        struktural_pieces: list[tuple[float, str]] = []
+        sprint_pieces: list[tuple[float, str]] = []
+        for j, ln in enumerate(ocr_lines):
+            if j in anchor["member_idxs"]:
+                continue
+            text = ln.text.strip()
+            if not text:
+                continue
+            x = _box_x_center(ln.box)
+            y = _box_y_center(ln.box)
+            if not (y_lo <= y <= y_hi):
+                continue
+            if x < pangkat_x - column_margin:
+                # NAMA side
+                if _RE_NAME_BLOCKLIST.match(text):
+                    continue
+                if _RE_BARE_ROW_NUMBER.match(text):
+                    continue
+                if not _is_plausible_name(text):
+                    continue
+                nama_pieces.append((y, text))
+            elif x > pangkat_x + column_margin:
+                # JABATAN side — split into STRUKTURAL vs DALAM SPRIN
+                # using the geometric column boundary detected above.
+                if _RE_NAME_BLOCKLIST.match(text):
+                    continue
+                if jabatan_split_x is not None and x > jabatan_split_x:
+                    sprint_pieces.append((y, text))
+                else:
+                    struktural_pieces.append((y, text))
+            # else: in PANGKAT column or column margin — skip
+
+        nama_pieces.sort(key=lambda p: p[0])
+        struktural_pieces.sort(key=lambda p: p[0])
+        sprint_pieces.sort(key=lambda p: p[0])
+
+        # Strip leading row number from the first nama piece (e.g. "1 F. GUNTUR"
+        # collapses to "F. GUNTUR" if the row marker happens to share a box).
+        if nama_pieces:
+            head = _RE_LEADING_ROW_NUMBER.sub("", nama_pieces[0][1]).strip()
+            nama_pieces[0] = (nama_pieces[0][0], head)
+
+        def _join(pieces: list[tuple[float, str]]) -> str | None:
+            text = " ".join(t for _, t in pieces if t).strip()
+            text = " ".join(text.split())
+            return text or None
+
+        rows.append(
+            PersonnelEntry(
+                no=None,
+                pangkat=anchor["pangkat"],
+                nrp=None,
+                nama=_join(nama_pieces),
+                jabatan_dinas=_join(struktural_pieces),
+                jabatan_sprint=_join(sprint_pieces),
+                keterangan=None,
+            )
+        )
+    return rows
+
+
+def is_low_quality(rows: list[PersonnelEntry]) -> bool:
+    """Heuristic: did PP-Structure produce useless rows?
+
+    A row is useful when it has at least pangkat OR nrp. If most rows have
+    only ``nama`` (or worse, nothing) the table extraction failed and the
+    caller should retry with the text-based fallback.
+    """
+    if not rows:
+        return True
+    useful = sum(1 for r in rows if r.pangkat or r.nrp)
+    # Require at least 30% of rows to carry rank/NRP signal. Below that we
+    # assume the column mapper degraded to "everything is nama" and prefer
+    # a fresh attempt.
+    return useful / max(1, len(rows)) < 0.3

src/ocr_sprint/pipeline/extract/regex_rules.py

@@ -53,19 +53,52 @@ _RE_TANGGAL_ID = re.compile(
     re.IGNORECASE,
 )
 
-# Satuan penerbit usually appears in the document letterhead, prefixed by
+# Polri letterhead pieces. The full letterhead spans multiple lines that are
-# KEPOLISIAN <NEGARA|DAERAH|RESORT|SEKTOR>.
+# often broken across separate OCR rows like:
-_RE_SATUAN = re.compile(
+#
-    r"KEPOLISIAN\s+(?:NEGARA\s+REPUBLIK\s+INDONESIA|DAERAH|RESOR(?:T)?|SEKTOR|RESORT)"
+#     KEPOLISIAN NEGARA REPUBLIK INDONESIA
-    r"[^\n]{0,80}",
+#     DAERAH JAWA BARAT
+#     RESOR CIMAHI
+#
+# We capture each individual level so we can reconstruct the most-specific
+# unit (RESOR / SEKTOR > DAERAH > NEGARA) — a downstream consumer cares
+# about *which* unit issued the sprint, not just that some Polri unit did.
+_RE_LEVEL_NEGARA = re.compile(
+    r"KEPOLISIAN\s+NEGARA\s+REPUBLIK\s+INDONESIA",
     re.IGNORECASE,
 )
+_RE_LEVEL_DAERAH = re.compile(
+    r"(?:KEPOLISIAN\s+)?DAERAH\s+([A-Z][A-Z .'/-]{1,60}?)(?:$|\s*\n)",
+    re.IGNORECASE | re.MULTILINE,
+)
+_RE_LEVEL_RESOR = re.compile(
+    r"(?:KEPOLISIAN\s+)?RESORT?\s+([A-Z][A-Z .'/-]{1,60}?)(?:$|\s*\n)",
+    re.IGNORECASE | re.MULTILINE,
+)
+_RE_LEVEL_SEKTOR = re.compile(
+    r"(?:KEPOLISIAN\s+)?SEKTOR\s+([A-Z][A-Z .'/-]{1,60}?)(?:$|\s*\n)",
+    re.IGNORECASE | re.MULTILINE,
+)
+_RE_LEVEL_MABES = re.compile(r"MABES\s+POLRI\b", re.IGNORECASE)
 
 # "Perihal : ...." up to end of line.
 _RE_PERIHAL = re.compile(r"PERIHAL\s*[:\-]\s*(.+)", re.IGNORECASE)
+# Many sprint docs (especially Polres-level) use 'Pertimbangan' as the
+# single-paragraph rationale block instead of (or alongside) 'Perihal'.
+# When `perihal` is missing we fall back to the first non-empty line under
+# 'Pertimbangan :' so the LLM doesn't have to guess and so a downstream
+# audit trail still has *something* in the perihal slot.
+_RE_PERTIMBANGAN_LABEL = re.compile(r"^\s*PERTIMBANGAN\b", re.IGNORECASE)
 
 # A dasar entry typically begins with a number and dot, e.g. "1. UU No. 2 Tahun 2002 ..."
 _RE_DASAR_ITEM = re.compile(r"^\s*(\d+)\s*[.)]\s*(.+)$")
+# OCR sometimes splits the number from its content across two lines:
+#     1.
+#      Undang-Undang Nomor 2 Tahun 2002 ...
+# We detect a bare-number line and merge with the next non-empty line.
+_RE_DASAR_BARE_NUMBER = re.compile(r"^\s*(\d+)\s*[.)]\s*$")
+# Generic 'untuk' bullet — same shape as a dasar item.
+_RE_UNTUK_ITEM = re.compile(r"^\s*(\d+)\s*[.)]\s*(.+)$")
 
 # Signatory NRP — Polri NRPs are 8 digits, civil servant NIPs are 18 digits.
 _RE_NRP = re.compile(r"\b(NRP|NIP)\s*[.:]?\s*(\d{8,20})\b", re.IGNORECASE)
@@ -99,54 +132,159 @@ def find_tanggal(text: str) -> date | None:
         return None
 
 
+def _clean_unit_tail(tail: str) -> str:
+    """Strip trailing punctuation/noise from the captured place name."""
+    return " ".join(tail.split()).strip(" .,;:'\"")
+
+
 def find_satuan(text: str) -> str | None:
-    """Return the first letterhead match (issuing unit), normalized."""
+    """Return the issuing unit, preferring the most-specific letterhead level.
-    match = _RE_SATUAN.search(text)
+
-    if not match:
+    Polri letterheads are hierarchical (Negara > Daerah > Resor/Sektor). The
-        return None
+    actual *issuing* unit is the deepest level present in the letterhead, not
-    return " ".join(match.group(0).split())
+    the topmost generic 'KEPOLISIAN NEGARA REPUBLIK INDONESIA' line. We scan
+    for each level independently and pick the most specific one available;
+    if only the generic Negara line is present we return that.
+
+    Examples
+    --------
+    >>> find_satuan("KEPOLISIAN NEGARA REPUBLIK INDONESIA\\n"
+    ...             "DAERAH JAWA BARAT\\nRESOR CIMAHI")
+    'KEPOLISIAN RESOR CIMAHI'
+    >>> find_satuan("KEPOLISIAN NEGARA REPUBLIK INDONESIA")
+    'KEPOLISIAN NEGARA REPUBLIK INDONESIA'
+    """
+    # We only look at the document head — letterheads always sit at the
+    # very top, and constraining the search prevents false positives from
+    # body text like '... Polres Cimahi ...' deep in a paragraph.
+    head = "\n".join(text.splitlines()[:25])
+
+    sektor = _RE_LEVEL_SEKTOR.search(head)
+    if sektor:
+        return f"KEPOLISIAN SEKTOR {_clean_unit_tail(sektor.group(1))}"
+    resor = _RE_LEVEL_RESOR.search(head)
+    if resor:
+        return f"KEPOLISIAN RESOR {_clean_unit_tail(resor.group(1))}"
+    daerah = _RE_LEVEL_DAERAH.search(head)
+    if daerah:
+        return f"KEPOLISIAN DAERAH {_clean_unit_tail(daerah.group(1))}"
+    if _RE_LEVEL_MABES.search(head):
+        return "MABES POLRI"
+    if _RE_LEVEL_NEGARA.search(head):
+        return "KEPOLISIAN NEGARA REPUBLIK INDONESIA"
+    return None
 
 
 def find_perihal(text: str) -> str | None:
-    """Return the first 'Perihal: ...' line, trimmed to that line only."""
+    """Return the first 'Perihal: ...' line, trimmed to that line only.
+
+    Falls back to the first non-empty line under a 'Pertimbangan' label
+    (a common variant in Polres-level surat sprint that doesn't have a
+    distinct 'Perihal' field). We deliberately keep this in regex-land
+    rather than deferring to the LLM because the LLM tends to hallucinate
+    perihal content from arbitrary paragraphs.
+    """
     for line in text.splitlines():
         m = _RE_PERIHAL.search(line)
         if m:
             return m.group(1).strip()
+
+    lines = text.splitlines()
+    for idx, line in enumerate(lines):
+        if _RE_PERTIMBANGAN_LABEL.match(line):
+            for follow in lines[idx + 1 : idx + 5]:
+                stripped = follow.strip(" :\t")
+                if stripped and stripped != ":":
+                    return stripped
+            break
     return None
 
 
+def _collect_numbered_section(
+    lines: list[str],
+    start_idx: int,
+    terminators: tuple[str, ...],
+) -> list[str]:
+    """Walk forward from ``start_idx`` collecting numbered list items.
+
+    Robust to OCR splitting the number marker onto its own line:
+        '1.'  ->   buffer ``pending_index=1``
+        next non-empty line starts the item body.
+
+    Continuation lines (non-empty, no leading number, after a started item)
+    are appended to the current item. Stops at any line whose uppercase form
+    starts with one of ``terminators``.
+    """
+    items: list[str] = []
+    pending_marker = False
+    blank_run = 0
+    for raw_line in lines[start_idx:]:
+        line = raw_line.strip()
+        upper = line.upper()
+        if any(upper.startswith(term) for term in terminators):
+            break
+        if not line:
+            blank_run += 1
+            # Two consecutive blank lines reliably mark the end of a section.
+            # A single blank line is tolerated because OCR sprinkles them.
+            if blank_run >= 2 and items and not pending_marker:
+                break
+            continue
+        blank_run = 0
+        bare = _RE_DASAR_BARE_NUMBER.match(line)
+        if bare:
+            pending_marker = True
+            continue
+        m = _RE_DASAR_ITEM.match(line)
+        if m:
+            items.append(m.group(2).strip())
+            pending_marker = False
+            continue
+        if pending_marker:
+            items.append(line)
+            pending_marker = False
+            continue
+        if items:
+            items[-1] = (items[-1] + " " + line).strip()
+    return items
+
+
 def find_dasar_list(text: str) -> list[str]:
     """Extract numbered 'Dasar' items from the text.
 
     Heuristic: locate a line containing 'DASAR' (Indonesian: "DASAR :") and
-    collect subsequent lines that start with a number. Stops at a blank line
+    delegate to ``_collect_numbered_section`` which handles three OCR
-    or a line beginning with another section header keyword.
+    artefacts:
+
+    1. Inline numbered items: ``"1. Undang-Undang ..."``.
+    2. Bare-number lines (the OCR engine puts the number alone on a line):
+       ``"1.\\n Undang-Undang ..."``.
+    3. Continuation lines (a line that is the wrapped tail of the previous
+       item gets appended back onto it).
     """
     lines = text.splitlines()
-    items: list[str] = []
-    in_dasar = False
     section_terminators = ("DIPERINTAHKAN", "UNTUK", "DASAR HUKUM", "PERIHAL")
-    for raw_line in lines:
+    for idx, raw_line in enumerate(lines):
-        line = raw_line.strip()
+        if re.match(r"^\s*DASAR\b", raw_line.strip(), re.IGNORECASE):
-        if not in_dasar:
+            return _collect_numbered_section(lines, idx + 1, section_terminators)
-            if re.match(r"^\s*DASAR\b", line, re.IGNORECASE):
+    return []
-                in_dasar = True
+
-            continue
+
-        if not line:
+def find_untuk_list(text: str) -> list[str]:
-            if items:
+    """Extract numbered 'Untuk' / 'DIPERINTAHKAN' bullets from the text.
-                break
+
-            continue
+    The 'Untuk' section follows 'DIPERINTAHKAN' / 'Kepada' and lists the
-        upper = line.upper()
+    tasks assigned to the personnel. Same OCR shape as Dasar, so we reuse
-        if any(upper.startswith(term) for term in section_terminators):
+    the collector but with different terminators.
-            break
+    """
-        m = _RE_DASAR_ITEM.match(line)
+    lines = text.splitlines()
-        if m:
+    # Stop conditions: 'Selesai' (boilerplate), 'Dikeluarkan di' (signature
-            items.append(m.group(2).strip())
+    # block), 'Tembusan' (carbon-copy section).
-        elif items:
+    terminators = ("SELESAI", "DIKELUARKAN", "TEMBUSAN", "PADA TANGGAL")
-            # continuation of the previous dasar item
+    for idx, raw_line in enumerate(lines):
-            items[-1] = (items[-1] + " " + line).strip()
+        if re.match(r"^\s*UNTUK\b", raw_line.strip(), re.IGNORECASE):
-    return items
+            return _collect_numbered_section(lines, idx + 1, terminators)
+    return []
 
 
 def find_signatory(text: str) -> Signatory:

src/ocr_sprint/pipeline/extract/validators.py

@@ -30,6 +30,13 @@ def validate_personnel_entry(entry: PersonnelEntry) -> list[ReviewFlag]:
         flags.append(ReviewFlag.INVALID_NRP)
     if entry.pangkat and not is_valid_pangkat(entry.pangkat):
         flags.append(ReviewFlag.UNKNOWN_PANGKAT)
+    # Identification of a personnel row requires at least pangkat OR nrp.
+    # A row carrying only a name is structurally incomplete - likely a
+    # mis-aligned table cell or a leaked tembusan/dasar fragment - and must
+    # be flagged for human review even though pangkat/nrp validation
+    # individually pass (because they're empty).
+    if not entry.pangkat and not entry.nrp:
+        flags.append(ReviewFlag.INCOMPLETE_PERSONNEL_ROW)
     return flags
 
 

src/ocr_sprint/pipeline/ocr.py

@@ -36,6 +36,73 @@ class OCRLine:
     box: tuple[tuple[float, float], ...]  # 4 (x, y) corner points
 
 
+def _line_y_center(line: OCRLine) -> float:
+    return sum(p[1] for p in line.box) / len(line.box)
+
+
+def _line_x_left(line: OCRLine) -> float:
+    return min(p[0] for p in line.box)
+
+
+def _line_height(line: OCRLine) -> float:
+    ys = [p[1] for p in line.box]
+    return max(ys) - min(ys)
+
+
+def sort_lines_by_layout(lines: list[OCRLine]) -> list[OCRLine]:
+    """Reorder lines into top-to-bottom, left-to-right reading order.
+
+    PaddleOCR's natural output order reflects detection order, not visual
+    layout. On dense tables (e.g. Polda Kalbar Akpol-panitia sprint) this
+    interleaves rows and columns — Paddle may emit a row's KET column
+    before its NAMA column, breaking every downstream extractor that
+    assumes top-to-bottom row order.
+
+    We rebuild reading order by:
+
+    1. Sorting by ``y_center``.
+    2. Grouping consecutive lines into row-bands when their ``y_center``
+       differs by less than half the median line height (so visually
+       same-row cells stay together even when their boxes don't perfectly
+       align).
+    3. Sorting each band left-to-right by ``x_left``.
+    """
+    if not lines:
+        return []
+
+    heights = [_line_height(ln) for ln in lines if _line_height(ln) > 0]
+    if not heights:
+        return list(lines)
+    median_height = sorted(heights)[len(heights) // 2]
+    band_threshold = max(1.0, median_height * 0.5)
+
+    by_y = sorted(lines, key=_line_y_center)
+    bands: list[list[OCRLine]] = []
+    current_band: list[OCRLine] = []
+    current_y: float | None = None
+    for ln in by_y:
+        y = _line_y_center(ln)
+        if current_y is None or abs(y - current_y) <= band_threshold:
+            current_band.append(ln)
+            # Track the band's running y-center as the mean of its
+            # members so a slowly-drifting set of cells doesn't split
+            # mid-row.
+            current_y = (
+                sum(_line_y_center(b) for b in current_band) / len(current_band)
+            )
+        else:
+            bands.append(current_band)
+            current_band = [ln]
+            current_y = y
+    if current_band:
+        bands.append(current_band)
+
+    ordered: list[OCRLine] = []
+    for band in bands:
+        ordered.extend(sorted(band, key=_line_x_left))
+    return ordered
+
+
 @dataclass(frozen=True)
 class OCRPage:
     """OCR output for a single page."""
@@ -44,8 +111,8 @@ class OCRPage:
 
     @property
     def text(self) -> str:
-        """Reconstruct page text by concatenating lines (order = paddle's output order)."""
+        """Reconstruct page text in visual reading order (top-to-bottom, left-to-right)."""
-        return "\n".join(line.text for line in self.lines)
+        return "\n".join(line.text for line in sort_lines_by_layout(self.lines))
 
     @property
     def mean_confidence(self) -> float:
@@ -55,9 +122,14 @@ class OCRPage:
 
 
 def _build_paddleocr() -> PaddleOCR:
+    s = get_settings()
+    if s.ocr_use_gpu:
+        from ocr_sprint.utils.gpu import configure_nvidia_dll_path
+
+        configure_nvidia_dll_path()
+
     from paddleocr import PaddleOCR
 
-    s = get_settings()
     kwargs: dict[str, object] = {
         "lang": s.ocr_lang,
         "use_angle_cls": True,
@@ -84,6 +156,19 @@ def get_ocr() -> PaddleOCR:
     return _instance
 
 
+def warmup() -> None:
+    """Eagerly initialize the PaddleOCR engine.
+
+    Call this during application startup so the first real request does not
+    pay the model-loading cost (~2-5s on CPU). Also prevents the process from
+    entering Disk-Sleep state (state D) mid-request when memory is tight,
+    because the OS has already paged in all model weights during startup.
+    """
+    _logger.info("paddleocr.warmup.start")
+    get_ocr()
+    _logger.info("paddleocr.warmup.done")
+
+
 def run_ocr(image: NDArrayU8) -> OCRPage:
     """Run OCR on a single BGR image and return a structured page result."""
     engine = get_ocr()

src/ocr_sprint/pipeline/orchestrator.py

@@ -19,7 +19,16 @@ from ocr_sprint.llm.extractor import llm_fill_header
 from ocr_sprint.pipeline.confidence import compute_confidence, route
 from ocr_sprint.pipeline.document_detect import DocumentDetectConfig, detect_and_correct
 from ocr_sprint.pipeline.extract.personnel import extract_personnel
-from ocr_sprint.pipeline.extract.regex_rules import extract_header, find_signatory
+from ocr_sprint.pipeline.extract.personnel_text import (
+    extract_personnel_from_ocr_lines,
+    extract_personnel_from_text,
+    is_low_quality,
+)
+from ocr_sprint.pipeline.extract.regex_rules import (
+    extract_header,
+    find_signatory,
+    find_untuk_list,
+)
 from ocr_sprint.pipeline.extract.validators import validate_extraction
 from ocr_sprint.pipeline.ingest import NDArrayU8, detect_source_kind, ingest
 from ocr_sprint.pipeline.ocr import OCRPage, run_ocr
@@ -112,6 +121,7 @@ def run_pipeline(content: bytes) -> PipelineOutput:
             header = merged
 
     personel: list[PersonnelEntry] = []
+    table_flags: list[ReviewFlag] = []
     if s.tables_enabled and cleaned_pages:
         all_tables: list[DetectedTable] = []
         for img in cleaned_pages:
@@ -126,14 +136,58 @@ def run_pipeline(content: bytes) -> PipelineOutput:
             personel_rows=len(personel),
         )
 
-    initial_flags: list[ReviewFlag] = list(llm_flags)
+    # Text-based fallback: PP-Structure can succeed structurally but emit
+    # rows with only ``nama`` populated (column mapper degraded), or fail to
+    # detect the table at all. In both cases the regex fallback that scans
+    # raw OCR for rank+NRP pairs produces a much more useful result. We
+    # always run it when the structured path is empty or low-quality, and
+    # raise a review flag so the operator knows the document didn't go
+    # through the preferred path.
+    if is_low_quality(personel):
+        fallback_rows = extract_personnel_from_text(full_text)
+        # If text-based fallback produced rows but they all lack NRP
+        # (Pass 3 territory), retry with the column-aware extractor that
+        # uses OCR bounding boxes. On dense tables (e.g. Polda Kalbar
+        # Akpol-panitia), text-only Pass 3 bleeds adjacent columns into
+        # nama/jabatan because lines are interleaved within each Y-band;
+        # the columnar variant restricts each field to its visual column.
+        text_only_no_nrp = bool(fallback_rows) and all(
+            r.nrp is None for r in fallback_rows
+        )
+        if (not fallback_rows) or text_only_no_nrp:
+            ocr_lines = [ln for page in ocr_pages for ln in page.lines]
+            columnar_rows = extract_personnel_from_ocr_lines(ocr_lines)
+            if columnar_rows and (
+                not fallback_rows or len(columnar_rows) >= len(fallback_rows)
+            ):
+                fallback_rows = columnar_rows
+        if fallback_rows:
+            personel = fallback_rows
+            # Pass 3 / columnar emit rows with nrp=None for sprint
+            # templates without an NRP column. Surface that with a
+            # distinct flag so operators know to expect missing NRPs by
+            # design rather than by OCR failure.
+            no_nrp = all(r.nrp is None for r in fallback_rows)
+            if no_nrp:
+                table_flags.append(ReviewFlag.PERSONNEL_TEXT_FALLBACK_NO_NRP)
+            else:
+                table_flags.append(ReviewFlag.PERSONNEL_TEXT_FALLBACK)
+            _logger.info(
+                "pipeline.personnel_text_fallback",
+                fallback_rows=len(fallback_rows),
+                no_nrp=no_nrp,
+            )
+
+    untuk_items = find_untuk_list(full_text)
+
+    initial_flags: list[ReviewFlag] = list(llm_flags) + list(table_flags)
     if mean_ocr_conf < _OCR_CONFIDENCE_FLAG_THRESHOLD:
         initial_flags.append(ReviewFlag.LOW_OCR_CONFIDENCE)
 
     result = ExtractionResult(
         header=header,
         personel=personel,
-        untuk=[],
+        untuk=untuk_items,
         ttd=ttd,
         raw_text=full_text,
         confidence=mean_ocr_conf,

src/ocr_sprint/pipeline/table.py

@@ -67,21 +67,43 @@ class DetectedTable:
 # ---------- PP-Structure singleton ----------
 
 
-def _build_pp_structure() -> PPStructure:
+def _create_pp_structure(
-    from paddleocr import PPStructure
+    pp_structure_cls: type[PPStructure], pp_lang: str, use_gpu: bool
-
+) -> PPStructure:
-    s = get_settings()
-    _logger.info("pp_structure.init", lang=s.ocr_lang, use_gpu=s.ocr_use_gpu)
     # layout=True so that PP-Structure also returns figure/text regions; we
     # filter to tables only afterwards. show_log=False to keep stdout clean.
-    return PPStructure(
+    return pp_structure_cls(
-        lang=s.ocr_lang,
+        lang=pp_lang,
-        use_gpu=s.ocr_use_gpu,
+        use_gpu=use_gpu,
         layout=True,
         show_log=False,
     )
 
 
+def _build_pp_structure() -> PPStructure:
+    s = get_settings()
+    if s.ocr_use_gpu:
+        from ocr_sprint.utils.gpu import configure_nvidia_dll_path
+
+        configure_nvidia_dll_path()
+
+    from paddleocr import PPStructure
+
+    # PPStructure layout models only support 'en' and 'ch', not 'latin'.
+    # Use 'en' for layout/table detection — it's language-agnostic (detects
+    # table structure, not text language). OCR within cells still works for
+    # Indonesian text because the recognition model handles Latin scripts.
+    pp_lang = "en" if s.ocr_lang not in ("en", "ch") else s.ocr_lang
+    _logger.info("pp_structure.init", lang=pp_lang, use_gpu=s.ocr_use_gpu)
+    try:
+        return _create_pp_structure(PPStructure, pp_lang, s.ocr_use_gpu)
+    except Exception as exc:
+        if not s.ocr_use_gpu:
+            raise
+        _logger.warning("pp_structure.gpu_init_failed_falling_back_cpu", error=str(exc))
+        return _create_pp_structure(PPStructure, pp_lang, False)
+
+
 def get_pp_structure() -> PPStructure:
     """Lazy, thread-safe singleton accessor for PP-Structure."""
     global _instance
@@ -92,6 +114,18 @@ def get_pp_structure() -> PPStructure:
     return _instance
 
 
+def warmup() -> None:
+    """Eagerly initialize the PP-Structure engine.
+
+    Call this during application startup so the first real request does not
+    pay the model-loading cost (~3-6s on CPU). Mirrors ocr.warmup() so the
+    lifespan handler can warm both engines in one place.
+    """
+    _logger.info("pp_structure.warmup.start")
+    get_pp_structure()
+    _logger.info("pp_structure.warmup.done")
+
+
 # ---------- table parsing ----------
 
 

src/ocr_sprint/schemas/extraction.py

@@ -21,6 +21,9 @@ class ReviewFlag(str, Enum):
     DATE_PARSE_FAILED = "date_parse_failed"
     LLM_FALLBACK = "llm_fallback"
     LLM_UNAVAILABLE = "llm_unavailable"
+    PERSONNEL_TEXT_FALLBACK = "personnel_text_fallback"
+    PERSONNEL_TEXT_FALLBACK_NO_NRP = "personnel_text_fallback_no_nrp"
+    INCOMPLETE_PERSONNEL_ROW = "incomplete_personnel_row"
 
 
 class Signatory(BaseModel):

src/ocr_sprint/utils/gpu.py

@@ -0,0 +1,57 @@
+"""GPU runtime helpers."""
+
+from __future__ import annotations
+
+import os
+from pathlib import Path
+
+_DLL_HANDLES: list[object] = []
+_CONFIGURED = False
+
+
+def configure_nvidia_dll_path() -> None:
+    """Expose NVIDIA wheel DLL directories to the Windows dynamic loader.
+
+    Paddle's Windows GPU wheels dynamically load CUDA/cuDNN DLLs by name. When
+    those DLLs come from Python packages such as ``nvidia-cudnn-cu11`` instead
+    of a system-wide CUDA Toolkit install, their ``bin`` folders are not on
+    ``PATH`` by default.
+    """
+    global _CONFIGURED
+    if _CONFIGURED or os.name != "nt":
+        return
+
+    package_names = ("nvidia.cudnn", "nvidia.cublas", "nvidia.cuda_nvrtc")
+    dll_dirs: list[Path] = []
+    for package_name in package_names:
+        try:
+            module = __import__(package_name, fromlist=["__file__"])
+        except Exception:
+            continue
+        module_file = getattr(module, "__file__", None)
+        if not module_file:
+            continue
+        dll_dir = Path(module_file).resolve().parent / "bin"
+        if dll_dir.is_dir():
+            dll_dirs.append(dll_dir)
+
+    if not dll_dirs:
+        _CONFIGURED = True
+        return
+
+    current_path_parts = os.environ.get("PATH", "").split(os.pathsep)
+    current_path_norm = {part.casefold() for part in current_path_parts if part}
+
+    prepend: list[str] = []
+    for dll_dir in dll_dirs:
+        dll_dir_str = str(dll_dir)
+        if dll_dir_str.casefold() not in current_path_norm:
+            prepend.append(dll_dir_str)
+        add_dll_directory = getattr(os, "add_dll_directory", None)
+        if add_dll_directory is not None:
+            _DLL_HANDLES.append(add_dll_directory(dll_dir_str))
+
+    if prepend:
+        os.environ["PATH"] = os.pathsep.join([*prepend, os.environ.get("PATH", "")])
+
+    _CONFIGURED = True

src/ocr_sprint/worker/celery_app.py

@@ -15,8 +15,12 @@ from __future__ import annotations
 import os
 
 from celery import Celery
+from celery.signals import worker_ready
 
 from ocr_sprint.config import get_settings
+from ocr_sprint.utils.logging import get_logger
+
+_logger = get_logger(__name__)
 
 
 def build_celery_app() -> Celery:
@@ -47,3 +51,32 @@ def build_celery_app() -> Celery:
 
 
 celery_app = build_celery_app()
+
+
+@worker_ready.connect
+def preload_ocr_models(sender: object, **kwargs: object) -> None:
+    """Warm up PaddleOCR and PP-Structure when the worker process is ready.
+
+    With ``--pool=solo`` the worker runs tasks in the *same* process that
+    receives this signal, so models loaded here are reused for every
+    subsequent task — no fork overhead, no duplicate model loading, and
+    RAM usage stays bounded (~1.5 GB instead of 1.5 GB × n_forks).
+    """
+    from ocr_sprint.config import get_settings as _gs
+    from ocr_sprint.pipeline import ocr as _ocr
+    from ocr_sprint.pipeline import table as _table
+
+    _logger.info("celery.worker.warmup.start")
+    s = _gs()
+    try:
+        _ocr.warmup()
+    except Exception as exc:
+        _logger.warning("celery.worker.paddleocr.warmup.failed", error=str(exc))
+
+    if s.tables_enabled:
+        try:
+            _table.warmup()
+        except Exception as exc:
+            _logger.warning("celery.worker.pp_structure.warmup.failed", error=str(exc))
+
+    _logger.info("celery.worker.warmup.done")

tests/unit/test_ground_truth_service.py

@@ -40,11 +40,19 @@ def _seed_approved_job_with_corrections(
             jid,
             status=DocumentStatus.NEEDS_REVIEW,
             confidence=0.8,
-            result=final_result
+            # ``is None`` (not truthiness) so callers can pass ``{}`` to
-            or {
+            # exercise the empty-dict edge case.
-                "header": {"nomor_sprint": "SPR/1/2025", "satuan_penerbit": "POLRES X"},
+            result=(
-                "personel": [{"pangkat": "AIPDA", "nrp": "77060000", "nama": "BUDI"}],
+                final_result
-            },
+                if final_result is not None
+                else {
+                    "header": {
+                        "nomor_sprint": "SPR/1/2025",
+                        "satuan_penerbit": "POLRES X",
+                    },
+                    "personel": [{"pangkat": "AIPDA", "nrp": "77060000", "nama": "BUDI"}],
+                }
+            ),
             review_flags=[],
         )
     if corrections:
@@ -197,6 +205,19 @@ def test_stats_counts_rollup_and_top_fields(db_ready: None) -> None:
     }
 
 
+def test_empty_dict_result_stays_consistent(db_ready: None) -> None:
+    """An empty-dict result (``{}``) is logically a valid snapshot — it
+    must round-trip as ``{}`` on *both* ``initial_result`` and
+    ``final_result``, not ``{}`` on one and ``None`` on the other.
+    """
+    _seed_approved_job_with_corrections(final_result={})
+    with session_scope() as session:
+        samples = list(iter_ground_truth_samples(session, GroundTruthFilters()))
+    assert len(samples) == 1
+    assert samples[0].initial_result == {}
+    assert samples[0].final_result == {}
+
+
 def test_serialize_is_valid_jsonl(db_ready: None) -> None:
     _seed_approved_job_with_corrections(corrections=[("header.perihal", "X", None)])
     with session_scope() as session:

tests/unit/test_ocr_layout.py

@@ -0,0 +1,75 @@
+"""Tests for OCR layout reordering.
+
+PaddleOCR emits text boxes in detection order, not visual reading order.
+On dense table layouts (Polda Kalbar Akpol-panitia regression) this
+interleaves columns within a row and breaks every downstream extractor
+that assumes top-to-bottom row order. ``sort_lines_by_layout`` rebuilds
+reading order from the bounding-box geometry.
+"""
+
+from __future__ import annotations
+
+from ocr_sprint.pipeline.ocr import OCRLine, OCRPage, sort_lines_by_layout
+
+
+def _box(x: float, y: float, w: float = 30, h: float = 15):
+    return ((x, y), (x + w, y), (x + w, y + h), (x, y + h))
+
+
+def _make(text: str, x: float, y: float) -> OCRLine:
+    return OCRLine(text=text, confidence=1.0, box=_box(x, y))
+
+
+class TestSortLinesByLayout:
+    def test_empty_returns_empty(self) -> None:
+        assert sort_lines_by_layout([]) == []
+
+    def test_already_sorted_is_stable(self) -> None:
+        lines = [_make("A", 10, 10), _make("B", 50, 10), _make("C", 10, 30)]
+        assert [ln.text for ln in sort_lines_by_layout(lines)] == ["A", "B", "C"]
+
+    def test_reorders_column_first_detection_to_row_first(self) -> None:
+        # Simulate a 2-row, 3-col table where Paddle returned cells
+        # column-first instead of row-first.
+        lines = [
+            _make("B1", 50, 10),
+            _make("B2", 50, 30),
+            _make("A1", 10, 10),
+            _make("A2", 10, 30),
+            _make("C1", 90, 10),
+            _make("C2", 90, 30),
+        ]
+        result = [ln.text for ln in sort_lines_by_layout(lines)]
+        assert result == ["A1", "B1", "C1", "A2", "B2", "C2"]
+
+    def test_groups_slightly_misaligned_cells_into_one_band(self) -> None:
+        # Real OCR boxes for a single visual row are rarely perfectly
+        # y-aligned; we still want them grouped.
+        lines = [
+            _make("LEFT", 10, 10),
+            _make("MID", 50, 12),  # 2px below LEFT — same row visually
+            _make("RIGHT", 90, 11),
+        ]
+        result = [ln.text for ln in sort_lines_by_layout(lines)]
+        assert result == ["LEFT", "MID", "RIGHT"]
+
+    def test_separates_rows_when_y_gap_exceeds_threshold(self) -> None:
+        # Lines with a y gap larger than ~½ line-height must NOT collapse
+        # into the same band.
+        lines = [
+            _make("ROW1A", 10, 10),
+            _make("ROW1B", 50, 10),
+            _make("ROW2A", 10, 30),  # gap of 20 vs height 15 → new band
+            _make("ROW2B", 50, 30),
+        ]
+        result = [ln.text for ln in sort_lines_by_layout(lines)]
+        assert result == ["ROW1A", "ROW1B", "ROW2A", "ROW2B"]
+
+    def test_ocrpage_text_uses_sorted_order(self) -> None:
+        lines = [
+            _make("RIGHT", 90, 10),
+            _make("LEFT", 10, 10),
+            _make("BOTTOM", 10, 30),
+        ]
+        page = OCRPage(lines=lines)
+        assert page.text == "LEFT\nRIGHT\nBOTTOM"

tests/unit/test_orchestrator_llm.py

@@ -169,3 +169,92 @@ def test_orchestrator_marks_unavailable_when_llm_returns_none(
     out = run_pipeline(b"%PDF-1.4\n%fake")
     assert ReviewFlag.LLM_UNAVAILABLE in out.result.review_flags
     assert ReviewFlag.LLM_FALLBACK not in out.result.review_flags
+
+
+def test_orchestrator_uses_text_fallback_when_pp_structure_yields_only_names(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    """When PP-Structure produces low-quality rows (e.g. only ``nama`` filled),
+    the orchestrator must run the text fallback against the raw OCR text and
+    raise the ``personnel_text_fallback`` flag.
+    """
+    monkeypatch.setenv("LLM_ENABLED", "false")
+    from ocr_sprint.config import get_settings
+
+    get_settings.cache_clear()
+
+    raw_text = (
+        "DAFTAR PERSONIL\n"
+        "1.\n"
+        "SRI WAHYUNI\n"
+        "AIPTU / 75070328\n"
+        "INTELKAM POLRES CIMAHI\n"
+        "2.\n"
+        "AGUNG LUKMAN\n"
+        "BRIPTU / 99030245\n"
+        "SAT INTELKAM\n"
+    )
+
+    # PP-Structure 'succeeded' but emitted name-only rows (the bug we saw on
+    # the real Polres Cimahi document).
+    from ocr_sprint.schemas.personnel import PersonnelEntry
+
+    pp_structure_low_quality = [
+        PersonnelEntry(nama="SRI WAHYUNI"),
+        PersonnelEntry(nama="AGUNG LUKMAN"),
+    ]
+    _stub_pipeline_stages(
+        monkeypatch,
+        raw_text=raw_text,
+        regex_header=HeaderFields(
+            nomor_sprint="Sprin/1/I/2025",
+            tanggal=date(2025, 1, 1),
+            satuan_penerbit="Polres Cimahi",
+            perihal="ok",
+            dasar=["UU 2/2002"],
+        ),
+    )
+    # Override extract_personnel to return the broken PP-Structure rows.
+    monkeypatch.setattr(orch_module, "extract_personnel", lambda _t: pp_structure_low_quality)
+
+    out = run_pipeline(b"%PDF-1.4\n%fake")
+    assert ReviewFlag.PERSONNEL_TEXT_FALLBACK in out.result.review_flags
+    # Fallback rows must carry pangkat + nrp (the whole point of the path).
+    assert all(r.pangkat and r.nrp for r in out.result.personel)
+    assert {r.pangkat for r in out.result.personel} == {"AIPTU", "BRIPTU"}
+
+
+def test_orchestrator_keeps_pp_structure_rows_when_quality_is_high(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    """Healthy PP-Structure output (rank+nrp present on most rows) must NOT
+    be replaced by the text fallback.
+    """
+    monkeypatch.setenv("LLM_ENABLED", "false")
+    from ocr_sprint.config import get_settings
+
+    get_settings.cache_clear()
+
+    from ocr_sprint.schemas.personnel import PersonnelEntry
+
+    healthy = [
+        PersonnelEntry(pangkat="AIPTU", nrp="11111111", nama="A"),
+        PersonnelEntry(pangkat="BRIPTU", nrp="22222222", nama="B"),
+        PersonnelEntry(pangkat="BRIPDA", nrp="33333333", nama="C"),
+    ]
+    _stub_pipeline_stages(
+        monkeypatch,
+        raw_text="ignored — should not be parsed",
+        regex_header=HeaderFields(
+            nomor_sprint="Sprin/1/I/2025",
+            tanggal=date(2025, 1, 1),
+            satuan_penerbit="Polres X",
+            perihal="ok",
+            dasar=["UU 2/2002"],
+        ),
+    )
+    monkeypatch.setattr(orch_module, "extract_personnel", lambda _t: healthy)
+
+    out = run_pipeline(b"%PDF-1.4\n%fake")
+    assert ReviewFlag.PERSONNEL_TEXT_FALLBACK not in out.result.review_flags
+    assert [r.nrp for r in out.result.personel] == ["11111111", "22222222", "33333333"]

tests/unit/test_personnel_text_fallback.py

@@ -0,0 +1,324 @@
+"""Tests for the text-based personnel fallback extractor.
+
+Driven by the real Polres Cimahi sprint document where PP-Structure
+produced 24 rows with only ``nama`` populated. The fallback should
+recover at least the rank + NRP for every row.
+"""
+
+from __future__ import annotations
+
+from ocr_sprint.pipeline.extract.personnel_text import (
+    extract_personnel_from_ocr_lines,
+    extract_personnel_from_text,
+    is_low_quality,
+)
+from ocr_sprint.pipeline.ocr import OCRLine
+from ocr_sprint.schemas.personnel import PersonnelEntry
+
+
+def _ocr_line(text: str, x: float, y: float, w: float = 80, h: float = 15) -> OCRLine:
+    box = ((x, y), (x + w, y), (x + w, y + h), (x, y + h))
+    return OCRLine(text=text, confidence=1.0, box=box)
+
+_CIMAHI_FIXTURE = """\
+DAFTAR PERSONIL SKCK POLRES DAN POLSEK JAJARAN POLRES CIMAHI TA 2024
+NO
+NAMA
+PANGKAT / NRP
+JABATAN
+KET
+BAUR SKCK SAT
+1.
+SRI WAHYUNI
+AIPTU / 75070328
+INTELKAM POLRES
+CIMAHI
+BA PELAKSANA SKCK
+2.
+CITRA DWI PUTRI R
+BRIPTU / 95070659
+ SAT INTELKAM
+POLRES CIMAHI
+BA PELAKSANA SKCK
+3.
+AGUNG LUKMAN AL
+BRIPTU / 99030245
+SAT INTELKAM
+POLRES CIMAHI
+BA POLSEK
+8.
+ARIEF SYAHRUL ZAMAN
+BRIGPOL /96030446
+MARGAASIH
+"""
+
+
+class TestExtractPersonnelFromText:
+    def test_extracts_rank_nrp_and_name(self) -> None:
+        rows = extract_personnel_from_text(_CIMAHI_FIXTURE)
+        assert len(rows) == 4
+        first = rows[0]
+        assert first.pangkat == "AIPTU"
+        assert first.nrp == "75070328"
+        assert first.nama == "SRI WAHYUNI"
+
+    def test_normalizes_brigpol_to_brigadir(self) -> None:
+        rows = extract_personnel_from_text(_CIMAHI_FIXTURE)
+        last = rows[-1]
+        # 'BRIGPOL' (no space) must canonicalize to 'BRIGADIR'.
+        assert last.pangkat == "BRIGADIR"
+        assert last.nrp == "96030446"
+        assert last.nama == "ARIEF SYAHRUL ZAMAN"
+
+    def test_skips_header_lines_as_names(self) -> None:
+        # No row should ever have a column-header word as nama.
+        rows = extract_personnel_from_text(_CIMAHI_FIXTURE)
+        names = [r.nama for r in rows]
+        for blocked in {"NAMA", "PANGKAT", "JABATAN", "KET", "DAFTAR"}:
+            assert blocked not in names
+
+    def test_jabatan_collected_from_following_lines(self) -> None:
+        rows = extract_personnel_from_text(_CIMAHI_FIXTURE)
+        assert rows[0].jabatan_dinas is not None
+        assert "INTELKAM" in rows[0].jabatan_dinas
+
+    def test_empty_text_returns_empty(self) -> None:
+        assert extract_personnel_from_text("") == []
+
+    def test_text_without_rank_nrp_pattern_returns_empty(self) -> None:
+        text = "Just a paragraph with no rank or NRP at all.\nAnother line."
+        assert extract_personnel_from_text(text) == []
+
+    def test_ignores_isolated_8digit_number_without_rank(self) -> None:
+        # NRP without a recognised rank token must not produce a row.
+        text = "Some line\n12345678\nanother line"
+        assert extract_personnel_from_text(text) == []
+
+    def test_rejects_unknown_rank_with_8digit_number(self) -> None:
+        # A "rank-shaped" word that isn't in the master list must not yield a row.
+        text = "Some line\nFAKERANK / 12345678\nanother line"
+        assert extract_personnel_from_text(text) == []
+
+    def test_does_not_drop_indonesian_names_starting_with_no_or_ket(self) -> None:
+        # Regression: 'NO' / 'KET' are legitimate column header tokens but
+        # also prefix common Indonesian names (KETUT, NOVA, NOOR). The
+        # blocklist must use word boundaries, not a raw startswith check.
+        text = (
+            "DAFTAR PERSONIL\n"
+            "1.\n"
+            "KETUT WARDANA\n"
+            "AIPTU / 11111111\n"
+            "JABATAN A\n"
+            "2.\n"
+            "NOVA SARI\n"
+            "BRIPTU / 22222222\n"
+            "JABATAN B\n"
+            "3.\n"
+            "NOOR HIDAYAT\n"
+            "BRIPDA / 33333333\n"
+            "JABATAN C\n"
+        )
+        rows = extract_personnel_from_text(text)
+        names = [r.nama for r in rows]
+        assert names == ["KETUT WARDANA", "NOVA SARI", "NOOR HIDAYAT"]
+
+    def test_extracts_multiple_rows_when_collapsed_to_one_line(self) -> None:
+        # Polres Banjar regression: when PaddleOCR merges several table
+        # rows onto a single OCR line, every rank+NRP pair on that line
+        # must still produce a separate row. Previously per-line
+        # ``re.search`` returned only the first match.
+        text = (
+            "DAFTAR NAMA INSTRUKTUR\n"
+            "1 CUCU JUHANA, A.K.S. KOMPOL 70100418 KABAGOPS "
+            "INSTRUKTUR LAT PRA OPS "
+            "HERU SAMSUL BAHRI, S.E., M.M., CPHR., CBA.AKP 77020049 "
+            "KASAT RESKRIM SDA "
+            "YAYAN SOPIANA, S.A.P., M.A.P., CPHR., CBA, CIAKP 84011113 "
+            "KASATINTELKAM POLRES BANJAR SDA\n"
+        )
+        rows = extract_personnel_from_text(text)
+        assert len(rows) == 3
+        assert [r.pangkat for r in rows] == ["KOMPOL", "AKP", "AKP"]
+        assert [r.nrp for r in rows] == ["70100418", "77020049", "84011113"]
+        assert rows[0].nama == "CUCU JUHANA, A.K.S."
+        assert rows[1].nama is not None and "HERU SAMSUL BAHRI" in rows[1].nama
+        assert rows[2].nama is not None and "YAYAN SOPIANA" in rows[2].nama
+
+    def test_extracts_multiple_rows_when_split_across_lines(self) -> None:
+        # Variant of the squished case where OCR produces one line per
+        # table row. Each row still ends up with multiple rank+NRP pairs
+        # never being on the same line, but verifies the finditer-based
+        # path doesn't regress this layout.
+        text = (
+            "1 CUCU JUHANA, A.K.S. KOMPOL 70100418 KABAGOPS\n"
+            "INSTRUKTUR LAT PRA OPS\n"
+            "HERU SAMSUL BAHRI, S.E., M.M., CPHR., CBA.AKP 77020049 KASAT RESKRIM\n"
+            "SDA\n"
+            "YAYAN SOPIANA, S.A.P., M.A.P., CPHR., CBA, CIAKP 84011113 KASATINTELKAM\n"
+            "POLRES BANJAR SDA\n"
+        )
+        rows = extract_personnel_from_text(text)
+        assert [r.pangkat for r in rows] == ["KOMPOL", "AKP", "AKP"]
+        assert [r.nrp for r in rows] == ["70100418", "77020049", "84011113"]
+        assert rows[0].nama == "CUCU JUHANA, A.K.S."
+
+    def test_extracts_rows_when_sprint_has_no_nrp_column(self) -> None:
+        # Polda Kalbar Akpol-panitia regression: sprint formats without
+        # an NRP column (panitia, undangan templates) must still extract
+        # rows via the rank-only Pass 3 path. Names span multiple OCR
+        # lines (narrow column), and the multi-token rank "KOMBES POL"
+        # is split across two lines.
+        text = (
+            "DAFTAR NAMA PANITIA\n"
+            "NO\nNAMA\nPANGKAT\nJABATAN\nSTRUKTURAL\nDALAM SPRIN\nKET\n"
+            "1\nF. GUNTUR\nSUNOTO, S.I.K.,\nM.H.\n"
+            "KOMBES\nPOL\n"
+            "KARO SDM\nPOLDA KALBAR\nKETUA\nPELAKSANA\n"
+            "2\nJUDA TRISNO\nTAMPUBOLON,\nS.H., S.I.K., M.H.\n"
+            "AKBP\n"
+            "KABAGDALPERS\nRO SDM\nPOLDA KALBAR\nSEKRETARIS\n"
+            "3\nPRAYITNO, S.H.,\nM.H.\n"
+            "KOMPOL\n"
+            "KASUBBAG DIAPERS\nANGGOTA\n"
+        )
+        rows = extract_personnel_from_text(text)
+        assert len(rows) == 3
+        assert [r.pangkat for r in rows] == ["KOMBES POL", "AKBP", "KOMPOL"]
+        # All Pass 3 rows have nrp=None by design.
+        assert all(r.nrp is None for r in rows)
+        assert rows[0].nama == "F. GUNTUR SUNOTO, S.I.K., M.H."
+        assert rows[1].nama == "JUDA TRISNO TAMPUBOLON, S.H., S.I.K., M.H."
+        assert rows[2].nama == "PRAYITNO, S.H., M.H."
+        assert rows[0].jabatan_dinas is not None and "KARO SDM" in rows[0].jabatan_dinas
+
+    def test_pass3_does_not_run_when_pass1_succeeds(self) -> None:
+        # If a sprint has NRPs (Pass 1 succeeds), Pass 3 must not fire
+        # and produce duplicate/contaminating rows.
+        text = (
+            "1\nSRI WAHYUNI\nAIPTU / 75070328\nBAUR SKCK\n"
+            "2\nCITRA DWI PUTRI\nBRIPTU / 95070659\nBA PELAKSANA\n"
+        )
+        rows = extract_personnel_from_text(text)
+        assert len(rows) == 2
+        assert all(r.nrp is not None for r in rows)
+
+    def test_still_blocks_bare_column_header_tokens(self) -> None:
+        # Word-boundary fix must still reject the actual column-header
+        # rows that motivated the blocklist in the first place.
+        text = "NO\nNAMA\nPANGKAT / NRP\nJABATAN\nKET\n1.\nREAL NAME\nAIPTU / 12345678\n"
+        rows = extract_personnel_from_text(text)
+        assert len(rows) == 1
+        assert rows[0].nama == "REAL NAME"
+
+
+class TestExtractPersonnelFromOcrLines:
+    """Column-aware Pass 3 — Polda Kalbar Akpol-panitia regression.
+
+    Verifies that bounding-box geometry preserves column boundaries on
+    dense tables where text-only Pass 3 bleeds adjacent columns into
+    nama/jabatan.
+    """
+
+    def _kalbar_lines(self) -> list[OCRLine]:
+        # Stylised Polda Kalbar layout: NO | NAMA | PANGKAT | STRUKTURAL | SPRIN
+        # X columns: 10, 100, 250, 380, 520. Each row may have multi-line cells.
+        return [
+            # Row 1 — KOMBES POL spans two stacked OCR boxes
+            _ocr_line("1", 10, 100),
+            _ocr_line("F. GUNTUR", 100, 100),
+            _ocr_line("SUNOTO, S.I.K.,", 100, 120),
+            _ocr_line("M.H.", 100, 140),
+            _ocr_line("KOMBES", 250, 100),
+            _ocr_line("POL", 250, 120),
+            _ocr_line("KARO SDM", 380, 100),
+            _ocr_line("POLDA KALBAR", 380, 120),
+            _ocr_line("KETUA", 520, 100),
+            _ocr_line("PELAKSANA", 520, 120),
+            # Row 2
+            _ocr_line("2", 10, 200),
+            _ocr_line("JUDA TRISNO", 100, 200),
+            _ocr_line("TAMPUBOLON,", 100, 220),
+            _ocr_line("S.H., S.I.K., M.H.", 100, 240),
+            _ocr_line("AKBP", 250, 200),
+            _ocr_line("KABAGDALPERS", 380, 200),
+            _ocr_line("RO SDM", 380, 220),
+            _ocr_line("POLDA KALBAR", 380, 240),
+            _ocr_line("SEKRETARIS", 520, 200),
+            # Row 9 — PNS PENATA TK I (multi-token rank stacked)
+            _ocr_line("9", 10, 500),
+            _ocr_line("FITRIANSYAH,", 100, 500),
+            _ocr_line("S.E.", 100, 520),
+            _ocr_line("PENATA", 250, 500),
+            _ocr_line("TK I", 250, 520),
+            _ocr_line("KAURKEU", 380, 500),
+            _ocr_line("RO SDM", 380, 520),
+            _ocr_line("POLDA KALBAR", 380, 540),
+            _ocr_line("BENDAHARA", 520, 500),
+        ]
+
+    def test_extracts_three_rows(self) -> None:
+        rows = extract_personnel_from_ocr_lines(self._kalbar_lines())
+        assert len(rows) == 3
+        assert [r.pangkat for r in rows] == ["KOMBES POL", "AKBP", "PENATA TK I"]
+
+    def test_nama_is_assembled_only_from_nama_column(self) -> None:
+        # Each row's nama must contain *all* its multi-line fragments
+        # and *only* its multi-line fragments — no bleed from struktural.
+        rows = extract_personnel_from_ocr_lines(self._kalbar_lines())
+        assert rows[0].nama == "F. GUNTUR SUNOTO, S.I.K., M.H."
+        assert rows[1].nama == "JUDA TRISNO TAMPUBOLON, S.H., S.I.K., M.H."
+        assert rows[2].nama == "FITRIANSYAH, S.E."
+
+    def test_jabatan_split_into_struktural_and_sprint(self) -> None:
+        # The geometric column boundary must split STRUKTURAL (jabatan_dinas)
+        # from DALAM SPRIN (jabatan_sprint).
+        rows = extract_personnel_from_ocr_lines(self._kalbar_lines())
+        assert rows[0].jabatan_dinas == "KARO SDM POLDA KALBAR"
+        assert rows[0].jabatan_sprint == "KETUA PELAKSANA"
+        assert rows[1].jabatan_dinas == "KABAGDALPERS RO SDM POLDA KALBAR"
+        assert rows[1].jabatan_sprint == "SEKRETARIS"
+
+    def test_returns_empty_when_no_rank_anchors(self) -> None:
+        lines = [
+            _ocr_line("DAFTAR NAMA", 100, 50),
+            _ocr_line("HEADER", 100, 100),
+        ]
+        assert extract_personnel_from_ocr_lines(lines) == []
+
+    def test_returns_empty_for_empty_input(self) -> None:
+        assert extract_personnel_from_ocr_lines([]) == []
+
+    def test_no_row_bleed_between_consecutive_rows(self) -> None:
+        # Row 1's last name fragment ("F. GUNTUR") sits BELOW its rank
+        # line but inside row 1's visual span. It must NOT leak into
+        # row 2's nama, which should start with "JUDA TRISNO".
+        rows = extract_personnel_from_ocr_lines(self._kalbar_lines())
+        assert rows[1].nama is not None
+        assert rows[1].nama.startswith("JUDA TRISNO")
+        assert "GUNTUR" not in rows[1].nama
+        assert "SUNOTO" not in rows[1].nama
+
+
+class TestIsLowQuality:
+    def test_empty_list_is_low_quality(self) -> None:
+        assert is_low_quality([]) is True
+
+    def test_all_rows_with_only_name_is_low_quality(self) -> None:
+        rows = [PersonnelEntry(nama=f"NAMA {i}") for i in range(10)]
+        assert is_low_quality(rows) is True
+
+    def test_majority_with_rank_nrp_is_high_quality(self) -> None:
+        rows = [
+            PersonnelEntry(nama=f"NAMA {i}", pangkat="AIPTU", nrp=f"{10000000 + i:08d}")
+            for i in range(10)
+        ]
+        assert is_low_quality(rows) is False
+
+    def test_borderline_30_percent_threshold(self) -> None:
+        # 3 useful out of 10 = exactly 0.3, treated as not-low-quality.
+        useful = [
+            PersonnelEntry(nama=f"NAMA {i}", pangkat="AIPTU", nrp=f"{10000000 + i:08d}")
+            for i in range(3)
+        ]
+        useless = [PersonnelEntry(nama=f"NAMA {i + 3}") for i in range(7)]
+        assert is_low_quality(useful + useless) is False

tests/unit/test_regex_rules.py

@@ -14,6 +14,7 @@ from ocr_sprint.pipeline.extract.regex_rules import (
     find_satuan,
     find_signatory,
     find_tanggal,
+    find_untuk_list,
 )
 
 
@@ -60,6 +61,36 @@ class TestSatuan:
         result = find_satuan("KEPOLISIAN NEGARA REPUBLIK INDONESIA")
         assert result is not None
 
+    def test_prefers_resor_over_negara_when_both_present(self) -> None:
+        # The Polri letterhead lists units hierarchically; the issuing unit
+        # is the deepest level, not the topmost generic "NEGARA" line.
+        text = (
+            "KEPOLISIAN NEGARA REPUBLIK INDONESIA\n"
+            "DAERAH JAWA BARAT\n"
+            "RESOR CIMAHI\n"
+            "SURAT PERINTAH\n"
+        )
+        result = find_satuan(text)
+        assert result == "KEPOLISIAN RESOR CIMAHI"
+
+    def test_prefers_sektor_over_resor(self) -> None:
+        text = (
+            "KEPOLISIAN NEGARA REPUBLIK INDONESIA\n"
+            "DAERAH JAWA BARAT\n"
+            "RESOR CIMAHI\n"
+            "SEKTOR PADALARANG\n"
+        )
+        result = find_satuan(text)
+        assert result == "KEPOLISIAN SEKTOR PADALARANG"
+
+    def test_handles_daerah_only(self) -> None:
+        text = "KEPOLISIAN NEGARA REPUBLIK INDONESIA\nDAERAH JAWA BARAT\n"
+        result = find_satuan(text)
+        assert result == "KEPOLISIAN DAERAH JAWA BARAT"
+
+    def test_returns_none_when_no_letterhead(self) -> None:
+        assert find_satuan("no police letterhead here") is None
+
 
 class TestPerihal:
     def test_extracts_perihal_line(self) -> None:
@@ -69,6 +100,25 @@ class TestPerihal:
     def test_returns_none_when_absent(self) -> None:
         assert find_perihal("no perihal field") is None
 
+    def test_falls_back_to_pertimbangan_block(self) -> None:
+        # Many Polres-level sprints use "Pertimbangan" instead of "Perihal".
+        # The fallback should pick up the first non-empty line under it.
+        text = (
+            "Pertimbangan\n"
+            "Bahwa dalam rangka mendukung kepentingan Dinas Polres Cimahi.\n"
+            "DASAR :\n"
+            "1. ...\n"
+        )
+        result = find_perihal(text)
+        assert result is not None
+        assert result.startswith("Bahwa dalam rangka mendukung")
+
+    def test_perihal_wins_over_pertimbangan_when_both_present(self) -> None:
+        # If the document has both a Perihal label AND a Pertimbangan
+        # paragraph, the explicit Perihal wins.
+        text = "Pertimbangan\nSome pertimbangan content.\nPERIHAL : The actual perihal.\n"
+        assert find_perihal(text) == "The actual perihal."
+
 
 class TestDasar:
     def test_numbered_list(self) -> None:
@@ -88,6 +138,57 @@ class TestDasar:
     def test_empty_when_section_missing(self) -> None:
         assert find_dasar_list("no dasar section") == []
 
+    def test_handles_bare_number_lines_split_by_ocr(self) -> None:
+        # OCR sometimes places the number marker on its own line and the
+        # body on the next non-empty line. The collector must merge them
+        # rather than dropping the body or appending it to the previous
+        # item (which the old implementation did).
+        text = (
+            "Dasar\n"
+            ":\n"
+            "1.\n"
+            " Undang - Undang Nomor 2 tahun 2002 tentang Kepolisian;\n"
+            "2. Peraturan Pemerintah Republik Indonesia No. 76 tahun 2020;\n"
+            "3.\n"
+            "Keterangan Catatan Kepolisian (SKCK);\n"
+            "4.\n"
+            "Pelayanan dilingkungan Badan Intelijen Keamanan Polri.\n"
+            "5. DIPA Petikan Satker Polres Cimahi.\n"
+            "DIPERINTAHKAN\n"
+        )
+        items = find_dasar_list(text)
+        assert len(items) == 5
+        assert items[0].startswith("Undang - Undang")
+        assert items[2].startswith("Keterangan Catatan")
+        assert items[3].startswith("Pelayanan dilingkungan")
+        assert items[4].startswith("DIPA")
+
+
+class TestUntuk:
+    def test_extracts_numbered_untuk_bullets(self) -> None:
+        text = (
+            "DIPERINTAHKAN\n"
+            "Kepada\n"
+            "Untuk\n"
+            "1.\n"
+            "melaksanakan tugas A;\n"
+            "2.\n"
+            "melaksanakan tugas B;\n"
+            "Selesai.\n"
+        )
+        items = find_untuk_list(text)
+        assert len(items) == 2
+        assert items[0] == "melaksanakan tugas A;"
+        assert items[1] == "melaksanakan tugas B;"
+
+    def test_returns_empty_when_section_missing(self) -> None:
+        assert find_untuk_list("no untuk section") == []
+
+    def test_stops_at_dikeluarkan(self) -> None:
+        text = "Untuk\n1. tugas A;\nDikeluarkan di Cimahi\n2. should not be captured\n"
+        items = find_untuk_list(text)
+        assert items == ["tugas A;"]
+
 
 class TestSignatory:
     def test_extracts_last_nrp(self) -> None:

tests/unit/test_table.py

@@ -2,8 +2,12 @@
 
 from __future__ import annotations
 
+import sys
+from types import ModuleType, SimpleNamespace
+
 import pytest
 
+from ocr_sprint.pipeline import table as table_module
 from ocr_sprint.pipeline.table import (
     DetectedTable,
     extract_tables_from_pp_result,
@@ -82,6 +86,34 @@ class TestDetectedTable:
         assert table.n_cols == 0
 
 
+class TestPpStructureInit:
+    def test_gpu_init_falls_back_to_cpu(self, monkeypatch: pytest.MonkeyPatch) -> None:
+        calls: list[dict[str, object]] = []
+
+        class FakePPStructure:
+            def __init__(self, **kwargs: object) -> None:
+                calls.append(kwargs)
+                if kwargs["use_gpu"]:
+                    raise RuntimeError("gpu init failed")
+
+        fake_paddleocr = ModuleType("paddleocr")
+        fake_paddleocr.PPStructure = FakePPStructure
+        monkeypatch.setitem(sys.modules, "paddleocr", fake_paddleocr)
+        monkeypatch.setattr(
+            table_module,
+            "get_settings",
+            lambda: SimpleNamespace(ocr_lang="latin", ocr_use_gpu=True),
+        )
+
+        engine = table_module._build_pp_structure()
+
+        assert isinstance(engine, FakePPStructure)
+        assert calls == [
+            {"lang": "en", "use_gpu": True, "layout": True, "show_log": False},
+            {"lang": "en", "use_gpu": False, "layout": True, "show_log": False},
+        ]
+
+
 @pytest.fixture
 def sample_personnel_table() -> DetectedTable:
     """Header + three personnel rows in a typical Polres-level format."""

tests/unit/test_validators.py

@@ -62,6 +62,20 @@ class TestPersonnelValidator:
         entry = PersonnelEntry(pangkat="Sersan Mayor", nrp="12345678", nama="Test")
         assert ReviewFlag.UNKNOWN_PANGKAT in validate_personnel_entry(entry)
 
+    def test_row_with_only_name_is_flagged_incomplete(self) -> None:
+        # A row that captured only `nama` (no pangkat AND no nrp) is the
+        # signature of mis-aligned table extraction. Must be flagged so
+        # the operator routes the document to needs_review.
+        entry = PersonnelEntry(nama="LEAKED FROM SOMEWHERE")
+        flags = validate_personnel_entry(entry)
+        assert ReviewFlag.INCOMPLETE_PERSONNEL_ROW in flags
+
+    def test_row_with_only_pangkat_is_not_flagged_incomplete(self) -> None:
+        # Having pangkat without NRP is suboptimal but still identifies a
+        # rank, so we don't raise the structural-incompleteness flag.
+        entry = PersonnelEntry(pangkat="AKP", nama="Test")
+        assert ReviewFlag.INCOMPLETE_PERSONNEL_ROW not in validate_personnel_entry(entry)
+
 
 class TestHeaderValidator:
     def test_complete_header_no_flags(self) -> None:

update.ps1

@@ -0,0 +1,214 @@
+#!/usr/bin/env pwsh
+# update.ps1 - One-command update & restart for ocr-sprint-service (local dev)
+
+param(
+    [ValidateSet("cpu", "gpu")]
+    [string] $OcrMode
+)
+
+$ErrorActionPreference = "Stop"
+
+$Port = 8000
+$ProjectRoot = $PSScriptRoot
+$VenvDir = Join-Path $ProjectRoot ".venv"
+$Python = Join-Path $VenvDir "Scripts\python.exe"
+
+function Invoke-Step {
+    param(
+        [Parameter(Mandatory = $true)]
+        [scriptblock] $Command,
+        [Parameter(Mandatory = $true)]
+        [string] $FailureMessage
+    )
+
+    & $Command
+    if ($LASTEXITCODE -ne 0) {
+        Write-Host "  $FailureMessage" -ForegroundColor Red
+        exit $LASTEXITCODE
+    }
+}
+
+function Get-DotEnvValue {
+    param(
+        [Parameter(Mandatory = $true)]
+        [string] $Name
+    )
+
+    $envFile = Join-Path $ProjectRoot ".env"
+    if (Test-Path $envFile) {
+        $line = Get-Content $envFile | Where-Object { $_ -match "^\s*$Name\s*=" } | Select-Object -Last 1
+        if ($line) {
+            return (($line -split "=", 2)[1] -split "\s+#", 2)[0].Trim()
+        }
+    }
+    return [Environment]::GetEnvironmentVariable($Name)
+}
+
+function Set-DotEnvValue {
+    param(
+        [Parameter(Mandatory = $true)]
+        [string] $Name,
+        [Parameter(Mandatory = $true)]
+        [string] $Value
+    )
+
+    $envFile = Join-Path $ProjectRoot ".env"
+    if (-not (Test-Path $envFile)) {
+        New-Item -Path $envFile -ItemType File | Out-Null
+    }
+
+    $lines = @(Get-Content $envFile)
+    $updated = $false
+    for ($i = 0; $i -lt $lines.Count; $i++) {
+        if ($lines[$i] -match "^\s*$Name\s*=") {
+            $comment = ""
+            if ($lines[$i] -match "(\s+#.*)$") {
+                $comment = $Matches[1]
+            }
+            $lines[$i] = "$Name=$Value$comment"
+            $updated = $true
+        }
+    }
+    if (-not $updated) {
+        $lines += "$Name=$Value"
+    }
+    Set-Content -Path $envFile -Value $lines
+}
+
+function Test-PythonPackage {
+    param(
+        [Parameter(Mandatory = $true)]
+        [string] $Name
+    )
+
+    & $Python -m pip show $Name *> $null
+    return $LASTEXITCODE -eq 0
+}
+
+function Add-NvidiaDllPaths {
+    $dllDirs = @(
+        (Join-Path $VenvDir "Lib\site-packages\nvidia\cudnn\bin"),
+        (Join-Path $VenvDir "Lib\site-packages\nvidia\cublas\bin"),
+        (Join-Path $VenvDir "Lib\site-packages\nvidia\cuda_nvrtc\bin")
+    )
+    foreach ($dir in $dllDirs) {
+        if ((Test-Path $dir) -and (($env:PATH -split ";") -notcontains $dir)) {
+            $env:PATH = "$dir;$env:PATH"
+        }
+    }
+}
+
+Set-Location $ProjectRoot
+
+if (-not (Test-Path $Python)) {
+    Write-Host "Virtualenv not found at $VenvDir. Creating one..." -ForegroundColor Yellow
+    $venvCreated = $false
+    $pythonLauncher = Get-Command py -ErrorAction SilentlyContinue
+    if ($pythonLauncher) {
+        foreach ($version in @("3.12", "3.11", "3.10")) {
+            & py "-$version" -m venv $VenvDir 2>$null
+            if ($LASTEXITCODE -eq 0) {
+                $venvCreated = $true
+                break
+            }
+        }
+    }
+    if (-not $venvCreated) {
+        $systemPython = Get-Command python -ErrorAction SilentlyContinue
+        if (-not $systemPython) {
+            Write-Host "  Python was not found. Install Python 3.10-3.12, then rerun this script." -ForegroundColor Red
+            exit 1
+        }
+        & python -m venv $VenvDir
+        $venvCreated = ($LASTEXITCODE -eq 0)
+    }
+    if (-not $venvCreated) {
+        Write-Host "  Failed to create virtualenv." -ForegroundColor Red
+        exit $LASTEXITCODE
+    }
+}
+
+$env:VIRTUAL_ENV = $VenvDir
+$env:PATH = "$(Join-Path $VenvDir 'Scripts');$env:PATH"
+
+if ($PSBoundParameters.ContainsKey("OcrMode")) {
+    $ocrUseGpuValue = if ($OcrMode -eq "gpu") { "true" } else { "false" }
+    Set-DotEnvValue "OCR_USE_GPU" $ocrUseGpuValue
+    $env:OCR_USE_GPU = $ocrUseGpuValue
+    Write-Host "OCR mode set to $($OcrMode.ToUpperInvariant()) and saved to .env." -ForegroundColor Green
+}
+
+# ── [1/5] Git pull ──────────────────────────────────────────────────────────
+Write-Host "`n[1/5] Pulling latest code..." -ForegroundColor Cyan
+Invoke-Step { git pull } "Git pull failed."
+
+# ── [2/5] Install/update dependencies ───────────────────────────────────────
+Write-Host "`n[2/5] Installing/updating dependencies..." -ForegroundColor Cyan
+Invoke-Step { & $Python -m pip install -e ".[dev]" -q } "Dependency install failed."
+
+$ocrUseGpu = (Get-DotEnvValue "OCR_USE_GPU")
+if ($ocrUseGpu -and $ocrUseGpu.ToLowerInvariant() -in @("1", "true", "yes", "on")) {
+    Write-Host "  GPU mode enabled; checking Paddle CUDA runtime..." -ForegroundColor Cyan
+    if (-not (Test-PythonPackage "paddlepaddle-gpu")) {
+        Invoke-Step {
+            & $Python -m pip install paddlepaddle-gpu==2.6.2 -i https://www.paddlepaddle.org.cn/packages/stable/cu118/ -q
+        } "Paddle GPU install failed."
+    }
+    if (-not (Test-PythonPackage "nvidia-cudnn-cu11")) {
+        Invoke-Step { & $Python -m pip install nvidia-cudnn-cu11==8.9.5.29 -q } "NVIDIA cuDNN install failed."
+    }
+    Add-NvidiaDllPaths
+} else {
+    Write-Host "  CPU mode enabled; checking Paddle CPU runtime..." -ForegroundColor Cyan
+    if (-not ((Test-PythonPackage "paddlepaddle") -or (Test-PythonPackage "paddlepaddle-gpu"))) {
+        Invoke-Step { & $Python -m pip install paddlepaddle==2.6.2 -q } "Paddle CPU install failed."
+    }
+}
+
+# ── [3/5] Database migration ─────────────────────────────────────────────────
+Write-Host "`n[3/5] Running database migrations..." -ForegroundColor Cyan
+& $Python -m alembic upgrade head
+if ($LASTEXITCODE -ne 0) {
+    Write-Host "  Migration conflict detected, stamping current state as head..." -ForegroundColor Yellow
+    Invoke-Step { & $Python -m alembic stamp head } "Alembic stamp failed."
+    Write-Host "  Retrying upgrade for any remaining new migrations..." -ForegroundColor Yellow
+    & $Python -m alembic upgrade head
+    if ($LASTEXITCODE -ne 0) {
+        Write-Host "  Migration still failed. Please check alembic manually." -ForegroundColor Red
+        exit 1
+    }
+}
+Write-Host "  Migrations OK." -ForegroundColor Green
+
+# ── [4/5] Free up port ───────────────────────────────────────────────────────
+Write-Host "`n[4/5] Checking port $Port..." -ForegroundColor Cyan
+
+# Use Get-NetTCPConnection for reliable port detection on Windows
+$connections = Get-NetTCPConnection -LocalPort $Port -State Listen -ErrorAction SilentlyContinue
+if ($connections) {
+    foreach ($conn in $connections) {
+        $procId = $conn.OwningProcess
+        $procName = (Get-Process -Id $procId -ErrorAction SilentlyContinue).Name
+        Write-Host "  Port $Port used by '$procName' (PID $procId), killing..." -ForegroundColor Yellow
+        Stop-Process -Id $procId -Force -ErrorAction SilentlyContinue
+    }
+    # Wait until port is actually released (max 5 seconds)
+    $waited = 0
+    do {
+        Start-Sleep -Milliseconds 500
+        $waited += 500
+        $still = Get-NetTCPConnection -LocalPort $Port -State Listen -ErrorAction SilentlyContinue
+    } while ($still -and $waited -lt 5000)
+
+    if ($still) {
+        Write-Host "  Port $Port still in use after waiting. Try a different port or restart manually." -ForegroundColor Red
+        exit 1
+    }
+    Write-Host "  Port $Port freed." -ForegroundColor Green
+} else {
+    Write-Host "  Port $Port is free." -ForegroundColor Green
+}
+
+# ── [5/5] Start dev server ───────────────────────────────────────────────────
+Write-Host "`n[5/5] Starting dev server on port $Port (Ctrl+C to stop)..." -ForegroundColor Cyan
+& $Python -m uvicorn ocr_sprint.main:app --reload --host 0.0.0.0 --port $Port