ezer-mishpati/legal-ai
1
0
Fork 0
Code Issues Pull Requests 2 Actions Packages Projects Releases Wiki Activity
129 Commits 8 Branches 1 Tag
437472be8578a66970582bb9741ba7ebc4ea69b8
Commit Graph

9 Commits

Author SHA1 Message Date
Chaim
22e819363e Flatten cases directory structure and unify paths 
- 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>
 2026-04-09 14:33:27 +00:00
Chaim
6aaca14e31 Replace Claude Vision OCR with Google Cloud Vision 
Benchmark results on Hebrew legal docs (case 1130-25):
- Google Vision: 1s/page, $0.001/page, high accuracy
- Claude Opus Vision: 90s/page, $0.05/page, poor accuracy
- PyMuPDF broken OCR layers now detected via quality check

Changes:
- extractor.py: Google Vision OCR with Hebrew language hint (300 DPI)
- extractor.py: text quality detection (word length, words-per-line, Hebrew ratio)
- extractor.py: Hebrew abbreviation quote fixer (15 known patterns)
- config.py: add GOOGLE_CLOUD_VISION_API_KEY, remove ANTHROPIC_API_KEY
- pyproject.toml: add google-cloud-vision, remove anthropic

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
 2026-04-08 20:17:58 +00:00
Chaim
bc72a83a71 Switch embedding model from voyage-3-large to voyage-law-2 
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>
 2026-04-08 19:05:58 +00:00
Chaim
5a8d5cac0a Add exports panel: versioned drafts, download, upload revisions, mark final 
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>
 2026-04-08 12:10:02 +00:00
Chaim
5fc52ce530 Switch to cases/{new,in-progress,completed}/ directory structure 
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>
 2026-04-04 10:45:47 +00:00
Chaim
e725f9ecd7 Fix claims parsing: truncated JSON recovery + chunking + compact output 
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>
 2026-04-03 16:04:34 +00:00
Chaim
e24e24dac5 Maximize context and output per Anthropic best practices 
Per official Anthropic documentation (April 2026):

Output tokens increased to match model capabilities:
- block-yod (discussion): 8K → 32K (Opus supports 128K)
- block-zayin (claims): 4K → 16K
- block-vav (background): 4K → 16K
- claims_extractor: 4K → 8K (fixes truncated JSON)
- qa_validator: 4K → 8K

Source documents sent in full (not truncated):
- Was: 3000 chars per doc, 15K total
- Now: full document text, no truncation
- Reduces hallucinations: "extract word-for-word quotes first"

Prompt structure follows long-context tips:
- Source documents placed FIRST (top of prompt)
- Instructions and query placed LAST
- "Queries at the end improve quality by up to 30%"

Extended thinking uses adaptive mode for Opus 4.6.
Streaming enabled for all requests > 21K tokens.

Unified JSON parsing via parse_llm_json() helper in config.py.
Applied to: classifier, claims_extractor, brainstorm, qa_validator,
learning_loop (5 files).

Also: extractor.py now supports .md files.

Sources:
- https://docs.anthropic.com/en/docs/build-with-claude/extended-thinking
- https://docs.anthropic.com/en/docs/build-with-claude/prompt-engineering/long-context-tips
- https://docs.anthropic.com/en/docs/minimizing-hallucinations
- https://docs.anthropic.com/en/docs/about-claude/models/overview

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
 2026-04-03 14:17:43 +00:00
Chaim
d9e5ef0f46 Add full decision writing pipeline: classify, extract, brainstorm, write, QA, export 
New services (11 files):
- classifier.py: auto doc-type classification + party identification (Claude Haiku)
- claims_extractor.py: claim extraction from pleadings (Claude Sonnet + regex)
- references_extractor.py: plan/case-law/legislation detection (regex)
- brainstorm.py: direction generation with 2-3 options (Claude Sonnet)
- block_writer.py: 12-block decision writer (template + Claude Sonnet/Opus)
- docx_exporter.py: DOCX export with David font, RTL, headings
- qa_validator.py: 6 QA checks with export blocking on critical failure
- learning_loop.py: draft vs final comparison + lesson extraction
- metrics.py: KPIs dashboard per case and global
- audit.py: action audit log
- cli.py: standalone CLI with 11 commands

Updated pipeline: extract → classify → chunk → embed → store → extract_references
New MCP tools: 29 total (was 16)
New DB tables: audit_log, decisions CRUD, claims CRUD
Config: Infisical support, external service allowlist

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
 2026-04-03 10:21:47 +00:00
Chaim
6f515dc2cb Initial commit: MCP server + web upload interface 
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
 2026-03-23 12:33:07 +00:00