adrian/OCR-SPRIN-SERVICE
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Fix personnel extraction + header bugs on real Polres Cimahi sprint

Browse Source 
This fixes 4 bugs found on a real Polres Cimahi SPRIN PDF:

1. satuan_penerbit captured the generic 'KEPOLISIAN NEGARA REPUBLIK
   INDONESIA' letterhead line instead of the most-specific issuing unit
   (e.g. RESOR CIMAHI / SEKTOR PADALARANG). Reworked find_satuan to
   scan for each level independently and return the deepest available.

2. find_dasar_list dropped numbered items when OCR put the marker on
   its own line ("1.\n Undang-Undang ..."). Refactored into
   _collect_numbered_section that buffers a bare-number line and uses
   the next non-empty line as the body. Also reused for the new
   find_untuk_list which extracts the previously-empty 'untuk' bullets.

3. find_perihal returned None for documents that use 'Pertimbangan'
   (very common in Polres-level sprint), forcing the LLM to guess.
   Added a regex fallback that picks up the first line under a
   'Pertimbangan' label so we keep extraction deterministic.

4. Personnel rows were emitted with only nama populated when
   PP-Structure detected a table but the column mapper degraded.
   Added a text-based fallback (extract_personnel_from_text) that
   scans raw OCR for <rank> + <8-digit NRP> patterns. Triggered when
   the PP-Structure result has fewer than 30% rank/NRP-bearing rows.
   Reviewed by raising the new PERSONNEL_TEXT_FALLBACK flag.

5. Validation now flags rows with neither pangkat nor nrp as
   INCOMPLETE_PERSONNEL_ROW, so the document routes to needs_review
   even when individual nrp/pangkat checks pass on empty values.

6. Added 'BRIGPOL' as a variant of BRIGADIR (seen in real scans).

Tests: 229 (was 203) — 26 new tests covering the regex fixes,
text-based personnel extractor, low-quality detector, validator
behaviour, and orchestrator wiring of the fallback path.

Co-Authored-By: adrian kuman firmansah <adriancuman@gmail.com>
This commit is contained in:
Devin AI
2026-04-26 05:35:42 +00:00
parent dce77e80e1
commit 58a2bf2648
11 changed files with 747 additions and 39 deletions
Download Patch File Download Diff File Expand all files Collapse all files

89
tests/unit/test_orchestrator_llm.py
View File

@@ -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"]

118
tests/unit/test_personnel_text_fallback.py Normal file
View File

@@ -0,0 +1,118 @@
"""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_text,
is_low_quality,
)
from ocr_sprint.schemas.personnel import PersonnelEntry
_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) == []
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

101
tests/unit/test_regex_rules.py
View File

@@ -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:

14
tests/unit/test_validators.py
View File

@@ -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: