leokariuki/cbk-policy-rag

RAG system over Central Bank of Kenya reports — retrieval evaluation portfolio project # CBK Policy RAG A retrieval-augmented generation system that answers questions about Kenyan monetary policy, banking supervision, and payments — grounded in the Central Bank of Kenya's published reports, with every answer citing the source document and page. **The headline is evaluation, not the chatbot.** This project demonstrates how retrieval and answer quality are *measured* and improved: baseline → change → metric moves. ## Status - [x] M0 — repo skeleton, config, secrets handling - [ ] M1 — ingest → clean → chunk → `corpus.jsonl` - [ ] M2 — embeddings + Chroma + dense retrieval - [ ] M3 — eval set (20–30 grounded Q/A pairs) - [ ] M4 — eval harness (recall@k, MRR, faithfulness) - [ ] M5 — hybrid retrieval + reranker, measured vs. baseline - [ ] M6 — Streamlit UI + writeup ## Setup ```bash python3 -m venv .venv source .venv/bin/activate pip install -r requirements.txt cp .env.example .env # then paste your free Groq key into .env ``` ## Constraints - Zero cost — local embeddings (sentence-transformers), local vector store (Chroma), free-tier LLM (Groq). - CBK PDFs and derived corpus are **never** committed (provenance/copyright). - No secrets in the repo — keys live in a gitignored `.env`.