Phase 5: hybrid LLM extraction (Ollama) for header gaps

Adds a small Ollama HTTP client (httpx-based, no extra runtime deps),
prompt builders, and a hybrid header extractor that runs *after* the
deterministic regex layer. The merger never overwrites a regex-filled
field — the LLM only fills gaps. If LLM_ENABLED=false (the default), or
the Ollama server is unreachable, the pipeline degrades gracefully:

  - LLM_ENABLED=false  ->  no LLM call at all, no flag.
  - LLM_ENABLED=true,
    header complete    ->  no LLM call.
  - LLM_ENABLED=true,
    header has gaps,
    LLM responded ok   ->  merge + LLM_FALLBACK flag (review hint).
  - LLM_ENABLED=true,
    header has gaps,
    LLM unavailable    ->  keep regex result + LLM_UNAVAILABLE flag.

Default model qwen2.5:1.5b on http://localhost:11434 — chosen for CPU
throughput (~5-15s per call) at acceptable accuracy. The LLM only fills
the *header* (nomor, tanggal, satuan, perihal, dasar). Personnel rows
stay with PP-Structure since that's more accurate and doesn't need LLM.

Tests:
 - test_llm_client.py: httpx MockTransport-driven tests for the wire
   format, error paths (HTTP 5xx, malformed JSON, missing envelope,
   ConnectError), and request shape.
 - test_llm_extractor.py: merge policy + None-on-unavailable behaviour.
 - test_orchestrator_llm.py: end-to-end orchestrator wiring with stubs
   for ingest/preprocess/OCR/table — verifies LLM is skipped when
   disabled, skipped when header is complete, called and flagged when
   gaps exist, and marked unavailable when the client returns None.

162 unit tests pass total (was 146).

Co-Authored-By: adrian kuman firmansah <adriancuman@gmail.com>

src/ocr_sprint/llm/__init__.py

@@ -0,0 +1,18 @@
+"""LLM-based extraction (Phase 5).
+
+The hybrid extractor first runs the deterministic regex layer and then —
+only for fields that came back missing or low-confidence — calls a local
+Ollama model with a Pydantic-typed prompt. Everything is gated by
+``LLM_ENABLED``; if the flag is off or the Ollama server is unreachable,
+the pipeline degrades gracefully back to the regex result.
+"""
+
+from ocr_sprint.llm.client import LLMUnavailableError, OllamaClient
+from ocr_sprint.llm.extractor import LLMHeaderResult, llm_fill_header
+
+__all__ = [
+    "LLMHeaderResult",
+    "LLMUnavailableError",
+    "OllamaClient",
+    "llm_fill_header",
+]

src/ocr_sprint/llm/client.py

@@ -0,0 +1,97 @@
+"""Ollama HTTP client.
+
+We deliberately avoid the ``ollama`` Python package — the wire format is a
+single ``POST /api/chat`` with ``format="json"`` and a system + user message,
+so a small ``httpx`` wrapper is enough. This keeps the runtime dependency
+footprint smaller and makes the mock-based unit tests trivial.
+"""
+
+from __future__ import annotations
+
+from typing import TypeVar
+
+import httpx
+from pydantic import BaseModel, ValidationError
+
+from ocr_sprint.config import get_settings
+from ocr_sprint.utils.logging import get_logger
+
+_logger = get_logger(__name__)
+
+T = TypeVar("T", bound=BaseModel)
+
+
+class LLMUnavailableError(RuntimeError):
+    """Raised when the Ollama server is unreachable, times out, or returns
+    a malformed payload. The pipeline catches this and falls back to the
+    regex-only result with a ``llm_fallback`` review flag.
+    """
+
+
+class OllamaClient:
+    """Tiny synchronous HTTP wrapper around the Ollama ``/api/chat`` endpoint.
+
+    Parameters
+    ----------
+    base_url:
+        Ollama server URL, e.g. ``http://localhost:11434``.
+    model:
+        Model tag to invoke (default ``qwen2.5:1.5b`` — chosen for CPU
+        latency at acceptable accuracy).
+    timeout_s:
+        Hard wall-clock timeout for a single request.
+    """
+
+    def __init__(
+        self,
+        base_url: str | None = None,
+        model: str | None = None,
+        timeout_s: int | None = None,
+    ) -> None:
+        s = get_settings()
+        self.base_url = (base_url or s.llm_base_url).rstrip("/")
+        self.model = model or s.llm_model
+        self.timeout_s = timeout_s if timeout_s is not None else s.llm_timeout_s
+
+    # ---------- public API ----------
+
+    def chat_json(self, system: str, user: str, schema_cls: type[T]) -> T:
+        """Run a single chat completion in JSON mode and validate the
+        response against ``schema_cls``. Raises ``LLMUnavailableError`` on
+        any transport / parse / validation failure so callers only have one
+        exception to handle.
+        """
+        payload = {
+            "model": self.model,
+            "stream": False,
+            "format": "json",
+            "messages": [
+                {"role": "system", "content": system},
+                {"role": "user", "content": user},
+            ],
+            # Keep determinism reasonable — we want extraction, not creativity.
+            "options": {"temperature": 0.0, "num_ctx": 4096},
+        }
+        url = f"{self.base_url}/api/chat"
+
+        try:
+            with httpx.Client(timeout=self.timeout_s) as client:
+                response = client.post(url, json=payload)
+                response.raise_for_status()
+                data = response.json()
+        except (httpx.HTTPError, ValueError) as exc:
+            _logger.warning("llm.transport_error", url=url, error=str(exc))
+            raise LLMUnavailableError(f"Ollama request failed: {exc}") from exc
+
+        # Ollama returns {"message": {"role": "assistant", "content": "<json>"}}.
+        try:
+            content = data["message"]["content"]
+        except (KeyError, TypeError) as exc:
+            _logger.warning("llm.bad_envelope", payload=data)
+            raise LLMUnavailableError(f"Ollama response missing message.content: {data!r}") from exc
+
+        try:
+            return schema_cls.model_validate_json(content)
+        except ValidationError as exc:
+            _logger.warning("llm.validation_error", error=str(exc), content=content[:400])
+            raise LLMUnavailableError(f"LLM JSON failed schema: {exc}") from exc

src/ocr_sprint/llm/extractor.py

@@ -0,0 +1,84 @@
+"""High-level LLM extractor.
+
+The job is *narrow*: take the raw OCR text plus the partial header that
+came back from the regex layer, and return an LLM-derived header that the
+caller can merge in. We never let the LLM populate the personnel table —
+PP-Structure is more accurate and cheaper for that.
+"""
+
+from __future__ import annotations
+
+from datetime import date
+
+from pydantic import BaseModel, Field
+
+from ocr_sprint.llm.client import LLMUnavailableError, OllamaClient
+from ocr_sprint.llm.prompts import SYSTEM_HEADER, build_user_prompt
+from ocr_sprint.schemas.extraction import HeaderFields
+from ocr_sprint.utils.logging import get_logger
+
+_logger = get_logger(__name__)
+
+
+class LLMHeaderResult(BaseModel):
+    """Schema we ask the model to fill. Mirrors ``HeaderFields`` but is
+    intentionally separate so we control exactly what the prompt and
+    validation surface look like — the public ``HeaderFields`` may grow
+    fields later that we don't want the LLM touching.
+    """
+
+    nomor_sprint: str | None = None
+    tanggal: date | None = None
+    satuan_penerbit: str | None = None
+    perihal: str | None = None
+    dasar: list[str] = Field(default_factory=list)
+
+
+def llm_fill_header(
+    raw_text: str,
+    regex_header: HeaderFields,
+    *,
+    client: OllamaClient | None = None,
+) -> HeaderFields | None:
+    """Run the LLM extractor and return a *merged* HeaderFields.
+
+    Returns ``None`` if the model is unavailable so the caller can decide
+    what to do (typically: keep the regex result and emit a fallback
+    review flag).
+    """
+    client = client or OllamaClient()
+
+    user = build_user_prompt(
+        raw_text=raw_text,
+        regex_partial=regex_header.model_dump(mode="json"),
+    )
+
+    try:
+        llm = client.chat_json(SYSTEM_HEADER, user, LLMHeaderResult)
+    except LLMUnavailableError as exc:
+        _logger.warning("llm.unavailable", error=str(exc))
+        return None
+
+    return _merge(regex_header, llm)
+
+
+def _merge(regex: HeaderFields, llm: LLMHeaderResult) -> HeaderFields:
+    """Merge LLM output into the regex result.
+
+    Policy: regex wins for any field it already filled. The LLM only fills
+    the *gaps*. This keeps deterministic / verifiable extractions for the
+    fields where regex is reliable and prevents the LLM from "correcting"
+    a value that happens to look unusual but is in fact correct.
+    """
+    merged = regex.model_copy(deep=True)
+    if merged.nomor_sprint is None and llm.nomor_sprint:
+        merged.nomor_sprint = llm.nomor_sprint
+    if merged.tanggal is None and llm.tanggal is not None:
+        merged.tanggal = llm.tanggal
+    if not merged.satuan_penerbit and llm.satuan_penerbit:
+        merged.satuan_penerbit = llm.satuan_penerbit
+    if not merged.perihal and llm.perihal:
+        merged.perihal = llm.perihal
+    if not merged.dasar and llm.dasar:
+        merged.dasar = list(llm.dasar)
+    return merged

src/ocr_sprint/llm/prompts.py

@@ -0,0 +1,48 @@
+"""Prompt builders for the LLM extractor.
+
+Kept in their own module so the prompts can be edited / version-tracked
+without touching the orchestration logic. We build prompts in Indonesian
+because the source documents are too — the model performs better when the
+field labels in the prompt match the OCR text it's being asked about.
+"""
+
+from __future__ import annotations
+
+SYSTEM_HEADER = (
+    "Anda adalah asisten ekstraksi data untuk dokumen Surat Perintah (Sprint) "
+    "Kepolisian Republik Indonesia (POLRI). Pengguna akan memberikan teks hasil "
+    "OCR sebuah surat sprint, dan Anda harus mengembalikan JSON yang sesuai "
+    "dengan skema yang diberikan.\n\n"
+    "Aturan keras:\n"
+    "1. Jangan mengarang. Jika sebuah field tidak terlihat di teks, kembalikan null.\n"
+    "2. Jangan menerjemahkan field. Output harus identik ejaannya dengan teks "
+    "sumber (kecuali normalisasi spasi/kapitalisasi yang jelas hasil OCR error).\n"
+    "3. Tanggal: kembalikan format ISO YYYY-MM-DD jika tanggal terlihat, "
+    "selain itu null.\n"
+    "4. Dasar hukum: array string berisi tiap butir, urut sesuai teks.\n"
+    "5. Jangan menambahkan field apa pun di luar skema. Output WAJIB JSON valid."
+)
+
+
+def build_user_prompt(raw_text: str, regex_partial: dict[str, object]) -> str:
+    """Construct the user message: OCR text + a hint about which fields the
+    deterministic regex layer already filled. Telling the LLM what we
+    *already have* keeps it from "creatively" overwriting good values.
+    """
+    known_fields = "\n".join(f"  - {k}: {v!r}" for k, v in sorted(regex_partial.items()) if v)
+    known_block = (
+        f"\nField yang sudah berhasil diekstrak dengan regex:\n{known_fields}\n"
+        if known_fields
+        else ""
+    )
+
+    return (
+        "Teks OCR:\n"
+        "----------\n"
+        f"{raw_text}\n"
+        "----------\n"
+        f"{known_block}"
+        "Tugas: kembalikan JSON dengan field nomor_sprint, tanggal (ISO date | null), "
+        "satuan_penerbit, perihal, dasar (array string). Hanya field yang terlihat — "
+        "yang tidak ada di teks isi null (atau array kosong untuk dasar)."
+    )

src/ocr_sprint/pipeline/orchestrator.py

@@ -15,6 +15,7 @@ from __future__ import annotations
 from dataclasses import dataclass
 
 from ocr_sprint.config import get_settings
+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
@@ -35,6 +36,18 @@ _logger = get_logger(__name__)
 _OCR_CONFIDENCE_FLAG_THRESHOLD = 0.80
 
 
+def _header_has_gaps(header: object) -> bool:
+    """True if any header field worth asking the LLM about is missing.
+
+    Using ``getattr`` so this stays decoupled from the exact attribute
+    names; the schema change cost was too large last time we hard-coded.
+    """
+    for field in ("nomor_sprint", "tanggal", "satuan_penerbit", "perihal"):
+        if not getattr(header, field, None):
+            return True
+    return not getattr(header, "dasar", None)
+
+
 @dataclass
 class PipelineOutput:
     """Bundle returned by the orchestrator."""
@@ -84,6 +97,20 @@ def run_pipeline(content: bytes) -> PipelineOutput:
     header = extract_header(full_text)
     ttd = find_signatory(full_text)
 
+    # Phase 5 — hybrid extraction. The regex layer is deterministic but
+    # brittle to layout variants between satuan; if any header field is
+    # still missing we ask the local LLM to fill the gaps. The merger
+    # never lets the LLM overwrite a field that regex already captured.
+    llm_flags: list[ReviewFlag] = []
+    if s.llm_enabled and _header_has_gaps(header):
+        merged = llm_fill_header(full_text, header)
+        if merged is None:
+            llm_flags.append(ReviewFlag.LLM_UNAVAILABLE)
+        else:
+            if merged.model_dump() != header.model_dump():
+                llm_flags.append(ReviewFlag.LLM_FALLBACK)
+            header = merged
+
     personel: list[PersonnelEntry] = []
     if s.tables_enabled and cleaned_pages:
         all_tables: list[DetectedTable] = []
@@ -99,7 +126,7 @@ def run_pipeline(content: bytes) -> PipelineOutput:
             personel_rows=len(personel),
         )
 
-    initial_flags: list[ReviewFlag] = []
+    initial_flags: list[ReviewFlag] = list(llm_flags)
     if mean_ocr_conf < _OCR_CONFIDENCE_FLAG_THRESHOLD:
         initial_flags.append(ReviewFlag.LOW_OCR_CONFIDENCE)
 

src/ocr_sprint/schemas/extraction.py

@@ -19,6 +19,8 @@ class ReviewFlag(str, Enum):
     UNKNOWN_PANGKAT = "unknown_pangkat"
     PERSONNEL_COUNT_MISMATCH = "personnel_count_mismatch"
     DATE_PARSE_FAILED = "date_parse_failed"
+    LLM_FALLBACK = "llm_fallback"
+    LLM_UNAVAILABLE = "llm_unavailable"
 
 
 class Signatory(BaseModel):