A retrieval-augmented question-answering assistant over public Indian financial documents (RBI, SEBI, mutual fund material), built around a real evaluation harness. It answers only from the documents, cites document and page, declines when the answer is not there, and routes risky questions to a human reviewer.
Status: Phase 0 of 8 (repository setup). No AI functionality exists yet. This README only describes what is built.
Everything runs on free tiers. See docs/DECISIONS.md.
- Docker Desktop
- uv (installs Python 3.12 for the project automatically)
- GNU make (Windows:
winget install ezwinports.make)
make setup # create .env from .env.example, install Python dependencies
make dev # start Postgres + pgvector, then the API on http://127.0.0.1:8000Check http://127.0.0.1:8000/health and the interactive API docs at http://127.0.0.1:8000/docs.
Other targets: make help.
make check # lint, strict type check, unit tests (same as CI)
make test-integration # tests against the real Postgres container| Folder | Contents |
|---|---|
api/ |
Python backend (FastAPI) |
infra/ |
Docker Compose for local services |
data/ |
Document manifest (Phase 1); downloaded PDFs are never committed |
eval/ |
Evaluation harness (Phase 2 and 5) |
web/ |
React frontend (Phase 7) |
docs/ |
Learning guide, decision records, build-in-public drafts |
NOT MEASURED. Evaluation starts in Phase 2.
Code is released under the MIT License. Source documents (RBI, SEBI, fund publications) are not part of this repository; they are downloaded at setup time and remain subject to their publishers' terms.