feat(retrieval): track page_number on text chunks for multimodal hybrid boost
All checks were successful
Build & Deploy / build-and-deploy (push) Successful in 6m33s

The legacy chunker did not track which PDF page each chunk came from.
Stored chunks had page_number=NULL, which blocked the multimodal
hybrid retriever's text+image boost — it joins (chunk, image) on
(document_id, page_number) and the join could never fire.

This change:

- extractor.extract_text now returns (text, page_count, page_offsets);
  page_offsets[i] is the start char offset of page (i+1) in the joined
  text. None for non-PDFs.
- chunker.chunk_document accepts an optional page_offsets and tags
  each chunk with the page that contains its first character (uses
  the existing chunker logic; pages assigned post-hoc by content
  search to keep the diff minimal).
- processor.process_document and precedent_library.ingest_precedent
  forward page_offsets through the chunker. New uploads now carry
  accurate page_number on every chunk.
- Other extract_text callers (tools/documents, tools/workflow,
  web/app.py) updated to unpack the third element (ignored).
- scripts/backfill_chunk_pages.py: per-case retrofit. Re-extracts each
  PDF (re-OCRs via Google Vision if needed, ~$0.0015/page), computes
  page_offsets, and updates page_number on every chunk by content
  search. Idempotent; --force re-runs on already-tagged docs.

Forward-only would leave the 419 image embeddings backfilled on
cases 8174-24 + 8137-24 unable to boost their corresponding text
chunks. The retrofit script closes that gap (cost ~$0.60).

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This commit is contained in:
2026-05-03 19:49:41 +00:00
parent 5724ed8e5b
commit 81ccf3a888
9 changed files with 301 additions and 18 deletions

View File

@@ -33,8 +33,15 @@ def chunk_document(
text: str,
chunk_size: int = config.CHUNK_SIZE_TOKENS,
overlap: int = config.CHUNK_OVERLAP_TOKENS,
page_offsets: list[int] | None = None,
) -> list[Chunk]:
"""Split a legal document into chunks, respecting section boundaries."""
"""Split a legal document into chunks, respecting section boundaries.
When ``page_offsets`` is supplied (from a PDF extraction), each chunk
is tagged with the page number of its first character — used by the
multimodal hybrid retriever to join (text chunk, image at same page)
and surface text+image matches.
"""
if not text.strip():
return []
@@ -52,9 +59,34 @@ def chunk_document(
))
idx += 1
if page_offsets:
_assign_pages(chunks, text, page_offsets)
return chunks
def _assign_pages(chunks: list[Chunk], text: str, page_offsets: list[int]) -> None:
"""Locate each chunk's first character in ``text`` and tag with the
page that contains that offset. Mutates chunks in-place.
Chunks have overlap so we search forward from a position slightly
past the previous chunk's start. Falls back to a global search if
the forward scan misses (rare — happens only when overlap is bigger
than the advance distance below).
"""
from legal_mcp.services.extractor import page_at_offset
pos = 0
for c in chunks:
idx = text.find(c.content, pos)
if idx < 0:
idx = text.find(c.content)
if idx < 0:
continue
c.page_number = page_at_offset(idx, page_offsets)
# advance past the chunk's halfway point — overlap is < 50% so
# the next chunk's starting point will be after this cursor.
pos = idx + max(1, len(c.content) // 2)
def _split_into_sections(text: str) -> list[tuple[str, str]]:
"""Split text into (section_type, text) pairs based on Hebrew headers."""
# Find all section headers and their positions

View File

@@ -120,12 +120,22 @@ def _fix_hebrew_quotes(text: str) -> str:
# ── Extraction ───────────────────────────────────────────────────
async def extract_text(file_path: str) -> tuple[str, int]:
# Separator used when joining per-page text. Constant so chunker /
# retrofit can reproduce the join when computing page offsets.
PAGE_SEPARATOR = "\n\n"
async def extract_text(file_path: str) -> tuple[str, int, list[int] | None]:
"""Extract text from a document file.
Returns:
Tuple of (extracted_text, page_count).
page_count is 0 for non-PDF files.
``(text, page_count, page_offsets)`` where:
- ``text``: concatenated extracted text
- ``page_count``: number of pages (0 for non-PDF)
- ``page_offsets``: ``page_offsets[i]`` = char start offset of
page (i+1) inside ``text``. ``None`` for non-PDFs (where the
notion of pages doesn't apply). Used by the chunker to assign
a ``page_number`` to each chunk.
"""
path = Path(file_path)
suffix = path.suffix.lower()
@@ -133,18 +143,34 @@ async def extract_text(file_path: str) -> tuple[str, int]:
if suffix == ".pdf":
return await _extract_pdf(path)
elif suffix == ".docx":
return _extract_docx(path), 0
return _extract_docx(path), 0, None
elif suffix == ".doc":
return _extract_doc(path), 0
return _extract_doc(path), 0, None
elif suffix == ".rtf":
return _extract_rtf(path), 0
return _extract_rtf(path), 0, None
elif suffix in (".txt", ".md"):
return path.read_text(encoding="utf-8"), 0
return path.read_text(encoding="utf-8"), 0, None
else:
raise ValueError(f"Unsupported file type: {suffix}")
async def _extract_pdf(path: Path) -> tuple[str, int]:
def _join_pages(pages_text: list[str]) -> tuple[str, list[int]]:
"""Join per-page text with PAGE_SEPARATOR while recording the start
offset of each page in the joined output."""
offsets: list[int] = []
parts: list[str] = []
cursor = 0
for i, pg in enumerate(pages_text):
offsets.append(cursor)
parts.append(pg)
cursor += len(pg)
if i < len(pages_text) - 1:
parts.append(PAGE_SEPARATOR)
cursor += len(PAGE_SEPARATOR)
return "".join(parts), offsets
async def _extract_pdf(path: Path) -> tuple[str, int, list[int]]:
"""Extract text from PDF.
Try direct text first, fall back to Google Cloud Vision for scanned
@@ -172,7 +198,27 @@ async def _extract_pdf(path: Path) -> tuple[str, int]:
pages_text.append(ocr_text)
doc.close()
return "\n\n".join(pages_text), page_count
joined, offsets = _join_pages(pages_text)
return joined, page_count, offsets
def page_at_offset(offset: int, page_offsets: list[int]) -> int:
"""Look up the page number containing a given char offset.
page_offsets[i] is the start of page (i+1) in the joined text;
a chunk starting at ``offset`` belongs to the highest-indexed page
whose start is ``<= offset``. Returns 1-based page number.
"""
if not page_offsets:
return 1
# Linear scan is fine — page_offsets is short (≤ ~200 for our PDFs).
page = 1
for i, start in enumerate(page_offsets):
if start <= offset:
page = i + 1
else:
break
return page
def _ocr_with_google_vision(image_bytes: bytes, page_num: int) -> str:

View File

@@ -127,7 +127,7 @@ async def ingest_precedent(
await progress("extracting", 15, "מחלץ טקסט מהקובץ")
try:
text, page_count = await extractor.extract_text(str(staged))
text, page_count, page_offsets = await extractor.extract_text(str(staged))
except Exception as e:
await progress("failed", 100, f"כשל בחילוץ טקסט: {e}")
raise
@@ -161,7 +161,7 @@ async def ingest_precedent(
try:
await progress("chunking", 40, f"מחלק את הטקסט ל-chunks ({page_count} עמ')")
chunks = chunker.chunk_document(text)
chunks = chunker.chunk_document(text, page_offsets=page_offsets)
if not chunks:
await db.set_case_law_extraction_status(case_law_id, "completed")
await db.set_case_law_halacha_status(case_law_id, "completed")

View File

@@ -32,7 +32,7 @@ async def process_document(document_id: UUID, case_id: UUID) -> dict:
try:
# Step 1: Extract text
logger.info("Extracting text from %s", doc["file_path"])
text, page_count = await extractor.extract_text(doc["file_path"])
text, page_count, page_offsets = await extractor.extract_text(doc["file_path"])
await db.update_document(
document_id,
@@ -70,9 +70,9 @@ async def process_document(document_id: UUID, case_id: UUID) -> dict:
except Exception as e:
logger.warning("Classification failed (non-fatal): %s", e)
# Step 2: Chunk
# Step 2: Chunk (page_offsets propagates page_number into chunks)
logger.info("Chunking document (%d chars)", len(text))
chunks = chunker.chunk_document(text)
chunks = chunker.chunk_document(text, page_offsets=page_offsets)
if not chunks:
await db.update_document(document_id, extraction_status="completed")

View File

@@ -144,7 +144,7 @@ async def document_upload_training(
shutil.copy2(str(source), str(dest))
# Extract text and strip Nevo preamble
text, page_count = await extractor.extract_text(str(dest))
text, page_count, _ = await extractor.extract_text(str(dest))
text = extractor.strip_nevo_preamble(text)
# Parse date

View File

@@ -308,7 +308,7 @@ async def ingest_final_version(
# Extract text from file if provided
if file_path and not final_text:
from legal_mcp.services import extractor
final_text, _ = await extractor.extract_text(file_path)
final_text, _, _ = await extractor.extract_text(file_path)
if not final_text:
return "לא סופק טקסט — יש לספק file_path או final_text."

View File

@@ -23,6 +23,7 @@
| `voyage_rerank_judge_poc.py` | python | POC #4 — voyage-3 vs rerank-2 vs context-3 על אהרון ברק, 18 שאילתות, claude-haiku-4-5 כ-judge. הכרעה: rerank-2 ניצח עם +9% mean@3 | בנצ'מרק חד-פעמי |
| `voyage_rerank_corpus_poc.py` | python | POC #5 — voyage-3 vs rerank-2 על קורפוס מלא (785 docs). הכרעה: +4.5% mean@3 כללי, +11.6% על P queries (practical) | בנצ'מרק חד-פעמי, אישר את שלב B |
| `multimodal_backfill.py` | python | Backfill voyage-multimodal-3 page embeddings על מסמכי תיקים קיימים. idempotent (skips by default), forces `MULTIMODAL_ENABLED=true` ל-run, רץ מהקונטיינר. שלב C — ראה `docs/voyage-upgrades-plan.md` | ידני per-case (`python multimodal_backfill.py 8174-24 8137-24`) |
| `backfill_chunk_pages.py` | python | Backfill `page_number` ב-`document_chunks` קיימים. legacy chunker לא tracked עמודים → `page_number=NULL` חוסם boost של multimodal hybrid (text+image join על אותו עמוד). re-extracts כל PDF (re-OCR אם צריך, ~$0.0015/page), מחשב page_offsets, ומעדכן chunks. idempotent | ידני per-case (`python backfill_chunk_pages.py 8174-24 8137-24`) |
## תיקיית `.archive/` — סקריפטים שהושלמו

View File

@@ -0,0 +1,204 @@
"""Backfill page_number on existing document_chunks.
Why this exists: the legacy chunker did not track which page each chunk
came from. After the page-tracking fix, new uploads carry page_number
correctly, but existing chunks have ``page_number=NULL`` in the DB.
That blocks the multimodal hybrid retriever's text+image boost (it
joins (chunk, image) on (document_id, page_number)).
What it does (per case):
1. List every document in the case
2. For each document with NULL page_number chunks:
a. Re-extract via extractor.extract_text (re-runs OCR if needed —
~$0.0015/page on Google Vision; idempotent on the DB side)
b. Compute page_offsets from the re-extracted text
c. For every chunk row (sorted by chunk_index), search its
content in the re-extracted text → look up page → UPDATE
3. Skip documents whose chunks already have non-null page_number
Idempotent: a second run with no --force is a no-op.
Designed to run from inside the FastAPI/MCP container (where /data
is mounted and Google Vision creds are present). Locally it requires
GOOGLE_CLOUD_VISION_API_KEY in ~/.env.
Usage:
docker exec -it <legal-ai-container> python /tmp/backfill_chunk_pages.py 8174-24 8137-24
"""
from __future__ import annotations
import argparse
import asyncio
import logging
import os
import sys
import time
from pathlib import Path
from uuid import UUID
def _setup_paths():
here = Path(__file__).resolve().parent
mcp_src = here.parent / "mcp-server" / "src"
if mcp_src.is_dir() and str(mcp_src) not in sys.path:
sys.path.insert(0, str(mcp_src))
_setup_paths()
from legal_mcp.services import db, extractor # noqa: E402
logging.basicConfig(
level=logging.INFO,
format="%(asctime)s [%(levelname)s] %(message)s",
)
logger = logging.getLogger("backfill_chunk_pages")
def _resolve_local_path(db_path: str) -> Path:
p = Path(db_path)
if p.is_file():
return p
if str(p).startswith("/data/"):
local = Path("/home/chaim/legal-ai") / Path(*p.parts[1:])
if local.is_file():
return local
return p
async def _backfill_document(
document_id: UUID,
title: str,
db_file_path: str,
force: bool,
) -> dict:
pool = await db.get_pool()
# Fetch chunks for this document
chunks = await pool.fetch(
"SELECT id, chunk_index, content, page_number FROM document_chunks "
"WHERE document_id = $1 ORDER BY chunk_index",
document_id,
)
if not chunks:
return {"status": "no_chunks"}
n_null = sum(1 for c in chunks if c["page_number"] is None)
if not force and n_null == 0:
logger.info(" skip (all %d chunks already tagged): %s", len(chunks), title)
return {"status": "skipped", "chunks": len(chunks)}
pdf_path = _resolve_local_path(db_file_path)
if not pdf_path.is_file():
logger.warning(" file missing: %s (%s)", pdf_path, title)
return {"status": "missing"}
if pdf_path.suffix.lower() != ".pdf":
return {"status": "not_pdf"}
logger.info(" re-extracting %s (%d chunks, %d need page)",
title, len(chunks), n_null)
t0 = time.time()
text, page_count, page_offsets = await extractor.extract_text(str(pdf_path))
elapsed = time.time() - t0
if not page_offsets:
return {"status": "no_offsets"}
# Walk chunks, find each in the re-extracted text, assign page
pos = 0
updated = 0
not_found = 0
for c in chunks:
content = c["content"]
if not content:
continue
idx = text.find(content, pos)
if idx < 0:
idx = text.find(content) # global fallback
if idx < 0:
not_found += 1
continue
page = extractor.page_at_offset(idx, page_offsets)
await pool.execute(
"UPDATE document_chunks SET page_number = $1 WHERE id = $2",
page, c["id"],
)
updated += 1
# advance roughly past midpoint — chunks have overlap
pos = idx + max(1, len(content) // 2)
logger.info(
" done in %.1fs: extracted %d pages, updated %d/%d chunks, "
"%d not found", elapsed, page_count, updated, len(chunks), not_found,
)
return {
"status": "ok",
"elapsed_sec": round(elapsed, 1),
"pages": page_count,
"chunks_total": len(chunks),
"chunks_updated": updated,
"chunks_not_found": not_found,
}
async def backfill_cases(case_numbers: list[str], force: bool) -> dict:
pool = await db.get_pool()
summary: dict = {}
for cn in case_numbers:
logger.info("=" * 60)
logger.info("Case %s", cn)
case = await db.get_case_by_number(cn)
if not case:
logger.warning("Case not found: %s", cn)
summary[cn] = {"status": "case_not_found"}
continue
case_id = UUID(str(case["id"]))
docs = await pool.fetch(
"SELECT id, title, file_path FROM documents WHERE case_id = $1 ORDER BY title",
case_id,
)
logger.info(" %d documents", len(docs))
per_doc: list[dict] = []
for d in docs:
r = await _backfill_document(
UUID(str(d["id"])), d["title"], d["file_path"], force,
)
per_doc.append({"document_id": str(d["id"]), "title": d["title"], **r})
summary[cn] = {
"documents_total": len(docs),
"ok": sum(1 for r in per_doc if r["status"] == "ok"),
"skipped": sum(1 for r in per_doc if r["status"] == "skipped"),
"missing": sum(1 for r in per_doc if r["status"] == "missing"),
"no_chunks": sum(1 for r in per_doc if r["status"] == "no_chunks"),
"no_offsets": sum(1 for r in per_doc if r["status"] == "no_offsets"),
"chunks_updated": sum(r.get("chunks_updated", 0) for r in per_doc),
"documents": per_doc,
}
return summary
def main():
parser = argparse.ArgumentParser(description="Backfill page_number on existing chunks")
parser.add_argument("cases", nargs="+", help="Case numbers (e.g. 8174-24 8137-24)")
parser.add_argument(
"--force", action="store_true",
help="Re-extract even if all chunks already have page_number (default: skip)",
)
args = parser.parse_args()
summary = asyncio.run(backfill_cases(args.cases, force=args.force))
print()
print("=" * 60)
print("SUMMARY")
print("=" * 60)
for cn, s in summary.items():
if s.get("status") == "case_not_found":
print(f" {cn}: NOT FOUND")
continue
print(
f" {cn}: {s['documents_total']} docs — "
f"ok {s['ok']}, skipped {s['skipped']}, missing {s['missing']}, "
f"chunks_updated {s['chunks_updated']}"
)
if __name__ == "__main__":
main()

View File

@@ -3500,7 +3500,7 @@ async def _process_training_document(task_id: str, source: Path, req: ClassifyRe
# Extract text
await _progress.set(task_id, {"status": "processing", "filename": req.filename, "step": "extracting"})
text, page_count = await extractor.extract_text(str(dest))
text, page_count, _ = await extractor.extract_text(str(dest))
# Parse date
d_date = None