README.md

metadata

title: legal-eye Hebrew Legal RAG
emoji: ⚖️
colorFrom: indigo
colorTo: blue
sdk: docker
app_port: 8000
pinned: false

legal-eye — Hebrew Legal RAG

Verbatim-from-precedent retrieval over Israeli case law and statutes.

No external LLM — answers are exact quotations from the corpus, anchored to a doctrine cluster (the leading case + its applications + its origins). The service refuses to run if ANTHROPIC_API_KEY or OPENAI_API_KEY are set in the environment together with TAU_RAG_AUTH_REQUIRED=true.

API surface (key endpoints)

• POST /v1/lawyer/ask — primary user endpoint. Accepts {question}, returns a structured response with confidence, arguments[], bundle, and disclaimer_he.
• POST /v1/hgraph/argument — bundle-only path. Faster and cheaper when the caller doesn't need the legacy synthesizer output.
• GET /v1/clusters?limit=N — list doctrine clusters.
• GET /v1/clusters/{cluster_id} — full cluster detail with members.
• GET /v1/system/dashboard — live corpus + retriever metrics.

All /v1/* endpoints require X-API-Key (set via the TAU_RAG_SEED_ADMIN_KEY Space secret).

Configuration

This Space loads the combined caselaw + kolzchut + statutes corpus (tau_rag/runtime/uploads/legal_eye_combined.jsonl) on first boot. Auto-loading parquet is intentionally disabled (TAU_RAG_AUTOLOAD_PARQUET=0) — the 2.6 GB raw judgments file does not fit comfortably on the free CPU tier.

Performance characteristics on free CPU (16 GB):

Phase	Time
First boot, retriever build (cold)	~2 min
First /v1/lawyer/ask (cluster build, lazy)	~30 s
Subsequent queries	2-5 s

The first query is slow because the doctrine cluster cache is built lazily. Once the index is persisted (after first boot), reboots reuse it.

Required Space secrets

Set these in Settings → Repository secrets before the Space first boots:

Name	How to generate	Purpose
TAU_RAG_SEED_ADMIN_KEY	openssl rand -hex 32	API key for /v1/* auth
TAU_RAG_HMAC_SECRET	openssl rand -hex 32	HMAC secret for signed URLs
TAU_RAG_CORS_ORIGINS	e.g. https://legal-eye.vercel.app	Frontend origin (CORS)

Disclaimer

⚠ This service is not legal advice. Output is a verbatim quotation from a court ruling or statute, surfaced via doctrine matching. Any legal decision must be reviewed by a licensed attorney.