חילוץ-המטא-דאטה של יומון (תג-מושג, כותרת-הלכה, מראה-מקום, תגיות מסיכום
עמוד-אחד) הוא משימה פשוטה בנפח גבוה — Sonnet הוא נקודת-האיזון מהירות/עלות,
בניגוד לחילוץ-הלכות שמצמיד Opus.
- config.DIGEST_EXTRACT_MODEL (env-tunable, ברירת-מחדל claude-sonnet-4-6).
- digest_metadata_extractor.extract(model=None) → ברירת-מחדל מה-config; קודם
לא צוין model → רץ על ברירת-המחדל של ה-CLI (Opus 4.8).
אומת: extract על יומון 5163 עם Sonnet החזיר תג-מושג/כותרת/מראה-מקום/תחום/
תגיות תקינים (~36s). claude_session נשאר local-only.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Completes #84 — surfaces the backend gating/prioritization (#84.1/#84.3, PR
#93) in the chair's review UI and adds near-duplicate clustering (#84.2).
Backend
- db.list_halachot gains `cluster` (#84.2): annotates each row with cluster_id +
cluster_size by unioning same-precedent halachot within HALACHA_CLUSTER_COSINE
(0.90, new config). Display-only — never merges/deletes. Pairwise is confined
to the returned set (cheap).
- GET /api/halachot exposes the `cluster` query param (default off).
Frontend (web-ui)
- Halacha type gains optional cluster_id / cluster_size (hand-written module; no
api:types regen needed — halachot aren't typed off the generated schema).
- useHalachotPending(opts): the default "clean" queue now fetches
exclude_low_quality + order_by_priority + cluster; needsFix:true returns the
flagged 'needs extraction fix' bucket (filtered client-side).
- HalachaReviewPanel: a "תור נקי / דורש תיקון-חילוץ" toggle (#84.1); near-dup
clusters collapse into ONE card showing "+N וריאנטים" with an expandable list,
and approve/reject/defer on a clustered card applies to all variants via the
batch endpoint (#84.2 + #84.4). Counts show true halacha totals (pendingTotal).
New flag labels added (application / near_duplicate / nevo_preamble_leak).
Verified:
- backend: list_halachot(cluster=True) on the live queue — algorithm correct
(groups related same-precedent rules at 0.78; none at the production 0.90
because dedup #82 already removed near-dups — the desired state).
- frontend: `tsc --noEmit` exits 0 (type-clean); no new lint errors (the one
lint error is pre-existing in training/learning-panel.tsx from #94). Local
Turbopack build can't run on the worktree node_modules symlink — CI builds in
a clean checkout.
Invariants: G1 (gate/cluster at source in SQL, not post-hoc); G2 (same
list_halachot path); §6 (flagged items routed to a visible bucket, not dropped).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Halacha-extraction quality (#81) and dedup-on-insert (#82) — engine changes
(pure + tested) plus measurement/ops tooling.
halacha_quality.py
- #81.4 application gate: is_fact_dependent() (high-precision "applied to THIS
case" deixis per the strict rubric §3/§27) + FLAG_APPLICATION. compute_quality_flags
now takes rule_type and flags rule_type=='application' OR fact-dependent —
blocking auto-approve (an illustration is not a generalizable holding).
- #82.3 lexical tail signal: jaccard_shingles / normalized_levenshtein /
lexical_near_duplicate + FLAG_NEAR_DUPLICATE, for the 0.83–0.93 cosine band.
halacha_extractor.py — pass rule_type to the flag computation; re-type a
binding-labeled fact-application to 'application' (mirrors non_decision→obiter).
db.py (store_halachot_for_chunk) — dedup now fetches the nearest same-precedent
neighbor once: cosine ≥ DEDUP → skip (unchanged); cosine in [BAND, DEDUP) with
high lexical overlap → FLAG_NEAR_DUPLICATE (review, not skip — never drop a
possibly-distinct principle unreviewed).
config.py — HALACHA_DEDUP_BAND_COSINE (0.83).
Scripts:
- scripts/halacha_goldset.py (#81.7) — export stratified sample for human
tagging; score validators (P/R/F1) against the tags. Backbone for #81.8.
- scripts/halacha_batch_reconcile.py (#82.7) — conservative cross-precedent
dedup (cosine ≥0.95), dry-run report only.
- scripts/calibrate_halacha_dedup.py (#82.1) — calibrate the lexical thresholds
against the 2026-06-03 cleanup gold-set.
Deferred (documented): #82.4 merge-provenance and #82.5 DB ON CONFLICT/UNIQUE
on normalized quote are NOT included — the current skip+flag behavior is safe,
whereas a UNIQUE on normalized_quote would fail on existing dups and a blind
merge risks losing provenance; they need their own chair-reviewed migration.
#82.6 over-merge guard is moot until merge lands. #81.6 full rhetorical-role
classifier deferred (section pre-filter + application flag cover the practical
case); #81.8 blocked on the human-tagged gold-set (harness now provided).
Verified:
- pytest tests/test_halacha_quality.py — 52 passed (14 new).
- calibrate: configured (0.55,0.70) → precision 1.0 (zero false-merge), recall
0.30 — correct profile for an auto-approve-blocking signal.
- goldset export: 15-row sample CSV. batch reconcile: 819 halachot → 5
cross-precedent candidate pairs.
Invariants: G1 (normalize at source — flag at insert, not at read); §6 (no
silent swallow — suspect items flagged to review, never dropped); G2 (no
parallel path — same store_halachot_for_chunk / compute_quality_flags).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
After a precedent finishes extracting, a claude_session pass folds facets of the
SAME legal question (below #82's dedup cosine — the שפר 14-vs-4 / 403-17→89
granularity gap) into one canonical; the rest are marked 'rejected' (reversible:
out of the active corpus AND the review queue, but recoverable). FOLD-ONLY —
never merges distinct legal questions, never invents.
- Engine: claude_session-as-judge (local CLI, zero cost), 'high' effort — folding
needs careful judgment. One pass per precedent, runs in _extract_impl once all
chunks are done (the prompt dedups within a chunk; this catches across chunks).
- Pure, unit-tested helpers in halacha_quality: CONSOLIDATE_SYSTEM,
build_consolidation_prompt, parse_fold_groups (fails SAFE → [] on any malformed
shape; drops <2-member groups; coerces/dedups indices).
- halacha_extractor._consolidate_precedent picks the canonical per group
(approved>pending, higher confidence, quote_verified, longer) and rejects the
rest via the existing update_halachot_batch (#84). Never rejects a canonical.
Fails OPEN on any error (no CLI / parse fail → 0 folds, data untouched).
- config: HALACHA_CONSOLIDATE_ENABLED/MODEL/EFFORT.
Verified: suite 176 passed (10 new); integration vs dev DB — a 2-facet group
folds to 1 canonical + 1 rejected (tagged), distinct rules untouched, claude
error → 0 folds (fail-open).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
#81.3 — a post-extraction validator that flags halachot whose rule_statement is
NOT entailed by its supporting_quote (the model over-reaching beyond its source).
- Engine: claude_session-as-judge (local CLI, zero API cost) per chaim's standing
preference — one batched judge call per chunk, NOT a hosted NLI model.
- Pure, unit-tested helpers in halacha_quality: NLI_SYSTEM, build_nli_prompt,
parse_nli_verdicts (fails OPEN — any shape/label ambiguity → 'entailed').
- halacha_extractor._nli_check wraps the call; fails OPEN on any error (e.g. no
CLI in the container) so a flaky judge never blocks a genuine halacha.
- Non-entailed (neutral/contradiction) → quality_flag 'nli_unsupported' which
blocks auto-approve (routes to pending_review) via the existing store gate.
- config: HALACHA_NLI_ENABLED/MODEL/EFFORT (effort 'low' — entailment is simple).
Verified: suite 166 passed (10 new); LIVE smoke test against the real claude CLI
returned ['entailed','neutral'] for a supported vs unsupported rule.
Also commits TaskMaster #86 (Nevo preamble/ratio: anti-contamination strip fix +
gold-set benchmark) capturing today's strip_nevo_preamble findings.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Bake the 2026-06-03 strict-cleanup rubric into the extraction pipeline so the
corpus stays clean at the source instead of accumulating duplicates, obiter
dicta, truncated quotes and thin restatements that clog the review queue.
#81 — quality gate:
- New pure module halacha_quality.py with unit-tested validators:
non-decision/obiter (Wambaugh markers), truncated-quote (mid-word cut),
thin-restatement (rule≈quote), quote-unverified.
- Validators run in halacha_extractor._process; a non-decision is re-typed
obiter; flags persist in new halachot.quality_flags column.
- Auto-approve now requires confidence>=threshold AND no quality flags;
flagged items route to pending_review regardless of confidence.
- Both extraction prompts hardened: reject undecided dicta, exclude
case-specific applications, require abstraction, forbid over-splitting.
#82 — dedup-on-insert (store_halachot_for_chunk):
- Within the same precedent, skip a halacha whose normalized supporting_quote
already exists, or whose rule-embedding has cosine>=HALACHA_DEDUP_COSINE
(0.93) against an already-stored one. Makes re-runs idempotent.
Migration: halachot.quality_flags TEXT[] (additive, idempotent ALTER).
Tests: 19 new unit tests; full suite 156 passed. Validated end-to-end against
dev DB (dedup skips dups, flag blocks auto-approve, re-run inserts 0).
Calibration: flags fire on only ~10% of current survivors (low false-positive).
Spec: docs/halacha-strict-rubric.md
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
xhigh is the quality sweet-spot for a single precedent but very slow at scale
(64-chunk case ≈ 20 min). Bulk queue-drains (process_pending over many
precedents) now use a lighter effort to cut wall-clock; interactive single
re-extraction keeps xhigh quality.
- config.HALACHA_BULK_EXTRACT_EFFORT (env, default 'high'; set 'medium' for max
speed, 'xhigh' to match single).
- extract()/_extract_impl()/_extract_chunk() take an `effort` override threaded
to claude_session.query_json; None falls back to HALACHA_EXTRACT_EFFORT (xhigh).
- process_pending_extractions(kind='halacha') passes the bulk effort; single
reextract_halachot keeps xhigh.
Verified end-to-end (mocked LLM): _extract_chunk(effort='medium') → query_json
effort='medium'; effort=None → 'xhigh' fallback. Closes the open item in #72.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
2026-05-31: opus-4-8 @ xhigh extraction + overlapping driver processes (agent
fallback retries each spawn an independent `python -c` driver; process_pending is
serial WITHIN a process but the box ran 4-5 drivers in parallel) → 12-16 concurrent
xhigh `claude -p` procs → load 69 → hard reboot.
Fix: halacha_extractor.extract() now takes a Postgres advisory lock
(pg_try_advisory_lock, key 'HALA') before any work. If another extraction (any
process/agent/driver — all share the legal-ai DB) holds it, the call returns
status='busy' and the precedent stays pending for the next drain. Guarantees ONE
extraction at a time ACROSS PROCESSES — an in-process Semaphore cannot (drivers
are separate OS processes). Core logic moved to _extract_impl (unchanged) under
the lock. CHUNK_CONCURRENCY now env-tunable (HALACHA_CHUNK_CONCURRENCY, default 3).
Verified: while a lock is held, extract() returns 'busy' with no LLM call; lock
releases cleanly and the next extraction proceeds. Tracks #72.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Five enhancements to the precedent retrieval stack:
* **#44 HNSW indexes** for precedent_chunks + halachot (replacing IVFFlat
lists=50). Build time ~3s combined. Better recall@10 with pgvector 0.8.2.
* **#45 Halacha sweep** — 96 pending halachot at conf>=0.78 promoted to
approved (1141 → 1237). Cluster at conf=0.78 spot-checked OK. Applied
via psql only — env HALACHA_AUTO_APPROVE_THRESHOLD unchanged (0.80).
* **#43 MMR diversity** — search_precedent_library_hybrid now caps at
``max_per_case_law=2`` (default). Prevents one precedent dominating
top-10 when many of its chunks/halachot rank high. New helper
``_diversify_by_case_law`` in hybrid_search.py.
* **#46 Dynamic halacha boost** — replaces the static ``score+=0.05``
with ``score+=confidence*0.06``. Calibrated so avg-confidence (~0.85)
stays at +0.05; high-conf halachot get a slight extra lift, low-conf
ones get less. Behaviour preserved at the mean.
* **#41 BM25/tsvector hybrid + RRF**. Schema V12 adds STORED tsvector
columns ``precedent_chunks.content_tsv`` and ``halachot.rule_tsv``
(using simple config — Postgres has no Hebrew stemmer) + GIN indexes.
New ``db.search_precedent_library_lexical`` mirrors the semantic
function with ts_rank_cd over plainto_tsquery. ``hybrid_search``
runs sem+lex in parallel and fuses via RRF before rerank. Toggle:
env ``BM25_HYBRID_ENABLED`` (default true), graceful fallback to
semantic-only on lexical failure.
#40 (VOYAGE_RERANK_ENABLED) was already true in Coolify env; no change.
#42 (Claude Haiku query expansion) deferred — latency + cost concerns
warrant a separate plan; the bm25 lexical leg already recovers most of
the exact-string recall #42 was meant to address.
Closes TaskMaster #41, #43-#46.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Cosine scores in voyage-3 (~0.4-0.5) and voyage-multimodal-3
(~0.2-0.25) live on different scales. The previous weighted-sum
merge let text always dominate — verified empirically: 0 image-only
hits across 7 queries on case 8174-24, image side contributed nothing.
RRF combines by *rank* in each list rather than raw score, robust
to scale differences. Per-item score:
rrf_score = text_weight / (k + text_rank)
+ image_weight / (k + image_rank)
A row that appears in both lists (joined on (id_field, page_number))
gets both terms — surfaced as match_type='text+image'.
After fix on 8174-24 (146 image rows): 2 image-only hits land in
top-5 across all 7 test queries, surfacing actual table/diagram/
signature pages (p12, p13 of שומת המשיבה for 'טבלת השוואת ערכי שומה',
p25 of שומת השגה for 'תרשים גוש וחלקה', etc).
On 8137-24 (273 image rows): 'חישוב היוון של דמי החכירה' goes from
0 baseline results → 5 hybrid results (3 text + 2 image), opening
recall on scanned content the OCR layer misses.
Default MULTIMODAL_TEXT_WEIGHT 0.65 → 0.5 (vanilla RRF) since the
prior 0.65 was tuned for raw cosine scales that no longer apply.
New env knob MULTIMODAL_RRF_K (default 60, standard literature).
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>
Halachot extracted by halacha_extractor with confidence >= 0.80 are now
inserted with review_status='approved' instead of 'pending_review' —
they appear in search_precedent_library immediately. Halachot below the
threshold still require manual chair approval.
Threshold tunable via env (HALACHA_AUTO_APPROVE_THRESHOLD), defaults to
0.80. Rationale: 89% of historical extractions (356/400) score 0.80+,
spot-checks confirmed quality, and the manual review backlog was the
single biggest reason rerank-2 was returning passages-only on
ההבחנה-style queries.
After this change + the one-time backfill UPDATE, search now returns
9/10 halachot for "ההבחנה בין השבחה לפיצויים" instead of 0 — and the
top-3 are exact-match rules, not adjacent passages.
Reviewer field records "auto-approved (confidence ≥ X.XX)" with the
threshold value at insert time, for traceability.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
- Remove cases/new|in-progress|completed subdivision (status managed in DB)
- Rename documents/original → documents/originals (consistent plural)
- Move exports from global data/exports/ into cases/{num}/exports/
- Add documents/research/ for case law and analysis files
- Update all agents, scripts, config, web API endpoints, and DB paths
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Benchmark on case 1130-25 (4 Hebrew legal docs, 8 queries) showed:
- voyage-law-2: avg top-1 score 0.5839 (+27% over voyage-3-large)
- voyage-4-large: avg top-1 score 0.4119 (worse than current)
- voyage-3-large: avg top-1 score 0.4589 (baseline)
voyage-law-2 costs ~4.6x more per run but delivers significantly
better retrieval quality for Hebrew legal text. Model is now
configurable via VOYAGE_MODEL env var.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Export DOCX now saves to data/exports/{case_number}/ with auto-versioning
(טיוטה-v1, v2...). The case view UI shows all drafts with download buttons,
allows uploading revised versions (עריכה-v1...), and marking a version as
final (copies to training corpus for style learning).
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Replace single CASES_DIR with find_case_dir() that searches across
all status directories. New cases created in cases/new/{number}/.
Config: CASES_BASE, CASES_NEW, CASES_IN_PROGRESS, CASES_COMPLETED
Docker: added -v /home/chaim/legal-ai/cases:/cases volume mount
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
config.py parse_llm_json: Added truncated JSON recovery. When Claude's
output is cut mid-JSON (common with long claim lists), the parser now:
- Finds the last complete JSON item (closing "}")
- Closes the array/object brackets
- Returns partial but valid results instead of None
Tested: recovers 2/3 items from truncated array, all cases pass.
claims_extractor.py:
- Prompt asks for compact output (150 words max per claim, group similar)
- Explicitly requests "no markdown, no explanations, JSON only"
- Long documents split into chunks at paragraph boundaries
- Each chunk processed separately, results merged
- max_tokens already at 8192
This fixes the recurring "0 claims" bug for committee responses and
permit applicant responses where the JSON was getting truncated.
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
