feat(retrieval): track page_number on text chunks for multimodal hybrid boost

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>

mcp-server/src/legal_mcp/services/chunker.py

@@ -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

mcp-server/src/legal_mcp/services/extractor.py

@@ -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:

mcp-server/src/legal_mcp/services/precedent_library.py

@@ -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")

mcp-server/src/legal_mcp/services/processor.py

@@ -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")

mcp-server/src/legal_mcp/tools/documents.py

@@ -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

mcp-server/src/legal_mcp/tools/workflow.py

@@ -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."