RegRadar

A retrieval-augmented generation (RAG) system for querying UK financial regulation — FCA Handbook, AML/KYC guidance, and PSD2/PSR text — with grounded, cited answers.

Status

🚧 In active development. Backend and frontend are both functional (Stages 1-12 of 16). Streaming, richer citations, deployment, and final docs remain.

What it does

RegRadar answers compliance questions against the FCA Handbook (SYSC sourcebook) with:

• Hybrid retrieval (dense embeddings + BM25 keyword search, fused via Reciprocal Rank Fusion)
• Cross-encoder reranking for precision
• LLM query rewriting (casual questions → precise regulatory phrasing)
• Citation-grounded generation — every claim traces back to a specific provision ID
• Explicit Rule vs. Guidance distinction (binding vs. non-binding FCA text), visually encoded in the UI
• Refuses to answer questions outside its ingested scope, rather than guessing

Stack

• Backend: FastAPI, Python 3.12
• Vector DB: Qdrant (hybrid dense + sparse search)
• Embeddings: OpenAI text-embedding-3-small
• Reranking: cross-encoder/ms-marco-MiniLM-L-6-v2 (sentence-transformers)
• Generation: OpenAI gpt-4o (query rewriting via gpt-4o-mini)
• Evaluation: RAGAS (faithfulness, answer relevancy), custom retrieval precision/recall
• Observability: Structured JSON logging, per-stage latency tracking, per-query cost tracking
• Frontend: React + TypeScript, Vite
• Deployment (planned): Docker, AWS EC2, GitHub Actions CI/CD

Architecture

See docs/ARCHITECTURE.md for the full system design, including the ingestion pipeline, query pipeline, and frontend structure.

See docs/eval_report.md for current evaluation metrics and known limitations.

Current data coverage

7 FCA Handbook SYSC sections (governance, systems/controls, financial crime, AML), 298 deduplicated chunks — a starting corpus, not the full Handbook.

Running locally

Backend:

# Start Qdrant
cd docker && docker compose up -d

# Install backend dependencies
cd ../backend
python3 -m venv venv && source venv/bin/activate
pip install -r requirements.txt

# Verify environment
python3 verify_setup.py

# Run ingestion (if starting fresh)
python3 ingestion/scrape_fca.py
python3 ingestion/clean_documents.py
python3 ingestion/chunk_semantic.py
python3 ingestion/generate_embeddings.py
python3 ingestion/upsert_to_qdrant.py
python3 retrieval/bm25_index.py

# Start the API
uvicorn api.main:app --reload

Frontend:

cd frontend
npm install
npm run dev

Visit http://localhost:5173. The frontend expects the backend running at http://localhost:8000 (configurable via frontend/.env, see .env.example).

Query the API directly:

curl -X POST http://localhost:8000/query \
  -H "Content-Type: application/json" \
  -d '{"query": "do i need to train my staff on money laundering?"}'

Evaluation

cd backend
python3 eval/run_eval.py       # run golden dataset through the pipeline
python3 eval/eval_report.py    # generate metrics report -> docs/eval_report.md

Frontend QA

See frontend/QA_CHECKLIST.md for the manual verification checklist covering all UI states (empty, loading, error, mobile, keyboard accessibility).

Why

Built to demonstrate production-grade RAG architecture — hybrid retrieval, reranking, query rewriting, citation-grounded generation, and a formal evaluation harness — not just embed-and-stuff-into-prompt. Designed around UK fintech compliance use cases (Revolut, Monzo, Citi-style risk/regulatory tooling).