CaseLogic — Source-Grounded Legal Research Assistant

Built from scratch in under 24 hours.

A source-grounded legal research assistant for personal-injury attorneys. Every claim is linked to a real public statute — no fabricated citations, no model-memory shortcuts.

Watch the demo →

For a short cheat sheet on adapters, extraction/retrieval/reasoning/verification, and the agent layer, see Architecture Q&A below (after the kickoff checklist).

Phased prep plan with technology guide: docs/plan.md.

Phase-1 MVP plan (CA Vehicle Code Harvester, query → UI): docs/phase1_plan.md. That doc describes the statute tables, hybrid retrieval, and /statutes API to build next; not all of it is implemented in the repo yet.

Current state

The repo is past blank scaffolding in a few areas, but most pipeline modules are still stubs.

Implemented

• FastAPI app (backend/main.py): CORS for http://localhost:3000, init_db on startup, /healthz.
• Status + ingest API (backend/api/):
  • GET /status — document counts + sample URLs from the documents table.
  • POST /ingest/search — Claude web_search_20250305 discovers URLs, then httpx fetches and persists rows (deduped by URL).
  • POST /ingest/url — direct URL fetch + persist.
• Persistence (backend/config.py, backend/db.py, backend/models.py): SQLite via SQLAlchemy; documents table only (no statutes / Phase-1 statute schema in code yet).
• Ingestion (backend/ingestion/): WebAdapter + ingest_search pipeline written end-to-end for web search → fetch → Document rows; base.py protocol; canlii / pdf adapters are still one-line stubs.
• Extraction types (backend/extraction/schemas.py): full Pydantic models for Variant A (PI case comparator) — PICaseFields, damages, liability, SourceSupport. extract.py / prompts.py are not wired to production extraction yet.
• Smoke check (backend/smoke_check.py): python -m backend.smoke_check verifies deps + ANTHROPIC_API_KEY.
• Frontend (frontend/): Next.js 15 + React 19 + Tailwind; three-column shell matching the baseline layout. Components still return null — no live API wiring.

Still stubs or missing

• Statute Harvester (Phase 1 per brief): Statute / StatuteFactor models, Chroma + FTS5 hybrid retrieval, GET /statutes/{id}, POST /statutes/search, GET /factors — specified in docs/phase1_plan.md; not present in backend/ yet (e.g. routes_statutes.py, retrieval/* implementations).
• Parsing, chunking, case-law extraction runtime, retrieval indexes, reasoning, verification — baseline modules exist as placeholder files.
• Agent layer — openclaw/tools.json / openclaw/agent_prompt.md not filled for statute tools.

Repo layout

.env.example              ANTHROPIC_API_KEY, DATABASE_URL, VECTOR_INDEX_PATH
requirements.txt          fastapi, anthropic, chromadb, sqlalchemy, httpx, ...
package.json              npm scripts that delegate to frontend/

backend/                  Python package — FastAPI app
  main.py                 app entry, routers: status + ingest
  config.py / db.py       runtime config + SQLAlchemy session
  models.py               documents table (statute tables = planned in phase1_plan)
  ingestion/              pipeline + adapters (web implemented; canlii/pdf stub)
  parsing/                stubs
  extraction/             schemas.py (PI case Pydantic); extract/prompts stub
  retrieval/              stubs (Chroma + hybrid planned)
  reasoning/              stubs
  verification/           stubs
  api/                    routes_ingest, routes_status (+ routes_search/answer/verify stub)
  smoke_check.py

openclaw/                 agent_prompt.md, tools.json — agent layer (minimal / empty)
frontend/                 Next.js app: app/, components/ (UI placeholders)
data/                     raw/, processed/, exports/, index/  (.gitkeep)
docs/                     plan.md, phase1_plan.md, architecture.md, demo_script.md, ...

Setup

python -m venv .venv
.venv\Scripts\activate    # Windows
pip install -r requirements.txt
cp .env.example .env      # then fill in ANTHROPIC_API_KEY (needed for /ingest/search)

Frontend (optional):

npm install --prefix frontend
npm run dev                 # from repo root: runs Next dev server (see package.json)

Run the backend

uvicorn backend.main:app --reload

Use /docs for interactive OpenAPI. Smallest working loop today: POST /ingest/url or /ingest/search → rows in SQLite + GET /status.

Next milestones (aligned with docs/phase1_plan.md)

1. Data: CA Vehicle Code ingest (e.g. leginfo) → statutes + statute_factors (or equivalent) with real official_url per row.
2. Retrieval: implement Chroma + SQLite FTS5 + RRF hybrid search; python -m backend.retrieval.build (or equivalent) after ingest.
3. API: statute lookup + search + factor listing; extend /status for statute counts and eval metadata.
4. UI: wire SearchPanel / ResultsPanel / SourceViewer / DatasetStatus to those endpoints.
5. Agent layer: tools pointing at the statute API for demos.

The checklist in docs/plan.md remains the broader baseline; docs/phase1_plan.md narrows Phase 1 to California-only query → result → UI on statutes.

Trust & safety

• Every claim traces back to a public URL + paragraph snippet
• Unsupported claims get marked, not hidden
• No legal advice; this is a research prototype

Disclaimer

Research prototype. Not legal advice. Results limited to indexed public sources.