#86.2 backfill + #86.3 benchmark, plus a #86.1 over-strip fix found en route.
extractor.py
- extract_nevo_ratio(): capture Nevo's מיני-רציו block (editorial holdings
summary) before it is stripped — a free professional gold-set (#86.3).
- _DECISION_START hardening (#86.2): the merged #86.1 regex over-stripped.
(a) פסק-דין headers are markdown-wrapped (**פסק דין**); the old anchor
required the keyword as the first line char with one separator, so it
missed the header and matched a citation 32K deep (עמ"נ 50567-07-21,
losing 45% of the body). Now tolerates leading markdown + 0-3 seps,
and the final-nun form (דין ן vs דינו נ).
(b) bare השופט/הנשיא matched CITATIONS ("השופט מ' חשין, פסקה 23"). The
authoring-judge line ends with a colon; we now require it.
ingest.py
- capture the ratio before stripping and store it on the row (best-effort,
non-fatal); also strip the text-upload path (was file-only).
db.py
- add case_law.nevo_ratio column (additive); allow it in update_case_law.
scripts/backfill_nevo_preamble.py (#86.2) — dry-run-by-default data migration:
finds historically-leaked rulings, captures ratio→nevo_ratio, rewrites
full_text (+content_hash), reindexes, and FLAGS (never deletes) halachot whose
quote lives in the removed preamble (review_status=pending_review +
nevo_preamble_leak flag). Safety guard: rows with keep%<--min-keep (60) are
excluded from --apply as suspected over-strip. --apply writes backup+manifest
to data/audit/ first. Chair-gated — NOT applied here.
scripts/nevo_ratio_benchmark.py (#86.3) — LLM-as-judge (local claude_session,
zero cost) measures recall/precision/granularity of our halachot vs the Nevo
ratio. Works pre- and post-backfill (reads nevo_ratio, falls back to full_text).
Verified:
- pytest tests/test_nevo_preamble.py — 12 passed (incl. citation/markdown
over-strip regressions).
- backfill dry-run: 19 leaked rulings, 27 contaminated halachot, all ≥75%
keep (the 32K over-strip is gone).
- benchmark on בג"ץ 1764/05: recall=0.875 precision=1.0 granularity=1.75x.
Invariants: G1 (normalize at source — strip/capture at ingest, not at read);
no silent swallow (contaminated halachot flagged + reported, not dropped);
data-migration is dry-run-default with backup+manifest, chair-gated.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
strip_nevo_preamble's _DECISION_START only matched ועדת-ערר openings (בפנינו /
הערר שבנדון / ...), so Nevo COURT judgments — exactly the ones carrying a
מיני-רציו — slipped through unstripped. The editorial mini-ratio then leaked into
the chunked body, risking that the halacha extractor reads Nevo's answer key
(contamination) and polluting the corpus. Proven on בג"ץ 1764/05: its full_text
still contained the מיני-רציו (unstripped).
Fix:
- Extend _DECISION_START with court-ruling openings: פסק-דין/פסק דין header and
the authoring-judge line (השופט/ת, כב' השופט, הנשיא, המשנה לנשיא). re.search
picks the earliest line-start match → the real opinion start, not the prose
ratio above it.
- Widen the Nevo-marker detection window 400→1500 chars so a long court/parties
header doesn't push חקיקה שאוזכרה:/מיני-רציו: out of range.
Verified on the real 1764/05 full_text: strips 2702 chars, body now starts at
'השופט ס' ג'ובראן:', מיני-רציו gone. Regression: ועדת-ערר openings still strip;
non-Nevo text untouched; markers-past-400 now detected. Suite 182 passed (6 new).
This is the anti-contamination prerequisite for the Nevo-ratio gold-set (#86.3/#81.7).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Legacy Hebrew .doc precedents (e.g. nevo.co.il CP1255 OLE2) can now be
uploaded directly through the precedent-library, missing-precedent, and
training upload paths — the frontend already advertised .doc but the
backend gate rejected it before reaching the extractor.
- web/app.py: add .doc to ALLOWED_EXTENSIONS (covers all paths that share
the set: precedent library, missing-precedent, training).
- Dockerfile: install libreoffice-writer-nogui (no X11/Java) so the
extractor's existing _extract_doc LibreOffice conversion works in the
Coolify container (was missing → would fail at runtime).
- extractor.py: isolate the LibreOffice user profile per call to avoid a
profile-lock failure on concurrent .doc conversions.
Verified in python:3.12-slim (prod base): .doc→.docx→text yields text
byte-identical to a native Word .docx save (103 paragraphs, 24,341 chars).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Born-digital Hebrew PDFs from legal software often encode gershayim (״)
as double-yod (יי), producing the same corruption patterns as OCR.
The fixer was only called after Google Cloud Vision OCR — digitally
created PDFs that passed quality checks received no correction.
Changes:
- Apply _fix_hebrew_quotes() in the direct extraction path
- Add 'בליימ' → 'בל"מ' (בקשה להארכת מועד — systematic corruption in 1017-03-26)
- Add 'תמייא' → 'תמ"א' (תכנית מתאר ארצית)
- Update docstring to reflect the broader scope
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
When Paperclip wakes the CEO and the model issues an mcp__legal-ai__*
call within ~10s of session init, Claude Code sometimes returns
"No such tool available" because the legal-ai MCP server hasn't
finished bringing up its tool catalog yet. Observed twice today on
CMPA precedent-extraction wakeups (sessions 9989fbaf and a9c61801);
the agent fell back to bash + .venv/bin/python and finished the work,
but the race needed fixing on the server side.
Three changes that close the window:
1. Lazy schema init (services/db.py + server.py)
`init_schema()` was awaited inside the FastMCP lifespan, blocking
the `initialize`/`tools/list` handshake until ~10 CREATE TABLE IF
NOT EXISTS statements ran. Under contention (two CEOs waking at
once for different companies) this stretched. Now the lifespan
returns immediately and `get_pool()` runs the schema migrations
exactly once on first DB access, guarded by an asyncio.Lock.
tools/list is answered in milliseconds regardless of DB state.
2. Lazy heavy imports
- services/embeddings.py: voyageai (~450ms) loaded only inside
_get_client()
- services/extractor.py: google.cloud.vision (~550ms) loaded only
inside _get_vision_client() and _ocr_with_google_vision()
These two were being imported at module top from
legal_mcp.tools.documents -> services.processor -> services.{
extractor,embeddings}, so the FastMCP server couldn't even start
responding until both finished. Cold start dropped from 2.7s to
1.17s end-to-end (init + tools/list response).
3. Agent-side warmup + retry guidance (.claude/agents/legal-ceo.md)
Even with a fast server, the model can still race on the very
first call. The precedent-extraction section now tells the CEO
to call workflow_status as a warmup probe and to retry after a
short sleep if it sees "No such tool available", before falling
back to the python bypass.
Also expanded the precedent-tool whitelists on the sub-agents that
delegate halacha/library work (commits 4a9a6b7 + 7ee90dc added the
tools to the MCP server but only the CEO got them in its allowed
list). Added to: legal-researcher (full extraction set), legal-analyst
(library_get/list + halacha review), legal-writer (library lookups +
halacha_review), legal-qa (library_get + halacha_review), and the two
that the CEO was already missing (halacha_review, halachot_pending).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
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>
Stage C: per-page image embeddings via voyage-multimodal-3 + hybrid
text+image search. Off by default; enable with MULTIMODAL_ENABLED=true.
- Schema V9: document_image_embeddings + precedent_image_embeddings
(vector(1024), page_number, image_thumbnail_path)
- extractor.render_pages_for_multimodal renders PDF pages at
MULTIMODAL_DPI (144) for embedding + JPEG thumbnails at
MULTIMODAL_THUMB_DPI (96) for UI preview, in one pass
- embeddings.embed_images calls voyage-multimodal-3 in 50-page batches
- services/hybrid_search.py orchestrator: rerank applied to text side
first (rerank-2 is text-only); image side cosine; weighted merge
with text_weight 0.65 (env-tunable); image-only pages surface as
match_type='image' so dense scanned content still appears
- processor.process_document and precedent_library.ingest_precedent
gated by flag — non-fatal on multimodal failure
- scripts/multimodal_backfill.py — idempotent per-case CLI to embed
existing documents without re-extracting text
Validated locally on a 5-page response brief: render 0.31s, embed 8.32s,
hybrid merge surfaces image rows correctly. Production rollout starts
with flag=false (no behavior change), then per-case A/B.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
- Add strip_nevo_preamble() to extractor.py — auto-removes Nevo database
headers (bibliography, legislation, mini-ratio) during training upload
- Add appeal_subtype column to style_patterns table — patterns are now
stored per subtype instead of globally mixed
- Update clear_style_patterns() to support subtype-scoped deletion
- Pass appeal_subtype through analyze_corpus → store → upsert pipeline
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Support ingestion of betterment levy (היטל השבחה) decisions into a
separate training corpus (CMPA). Key changes:
- Add .doc file extraction via LibreOffice conversion in extractor
- Add practice_area/appeal_subtype columns to style_corpus table
- Route training files to cmp/ or cmpa/ subdirs based on appeal subtype
- Fix derive_subtype to handle ARAR-YY-NNNN format (was matching year digit)
- Expose practice_area/appeal_subtype params in MCP upload_training tool
- Add appeal_subtype filter to analyze_style for per-type style analysis
- Update betterment levy methodology in lessons.py: checklist (from generic
to corpus-based), opening/closing strategies, and discussion rules
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Ezer Mishpati - AI legal decision drafting system with:
- MCP server (FastMCP) with document processing pipeline
- Web upload interface (FastAPI) for file upload and classification
- pgvector-based semantic search
- Hebrew legal document chunking and embedding
