TrustRAG Finance
Evaluation-driven financial RAG — hybrid retrieval, verified citations, abstention, and a hash-chained audit ledger
TrustRAG Finance is an evaluation-driven, read-only wealth-research assistant. It retrieves public financial documents, generates fully-cited answers, verifies every claim against the source passages, scores confidence from system signals rather than the model’s self-report, routes uncertain answers to human review, and measures itself against a golden dataset. The failure it exists to prevent is a confident wrong answer — an abstention is acceptable.
Why It Is Not a Normal RAG Chatbot
| Most RAG demos | TrustRAG Finance |
|---|---|
| Trust fluent output | Deterministic citation verification on every claim |
| Model self-rated confidence | Confidence from retrieval agreement, groundedness, and citation validity |
| Always answer | Abstains when evidence is insufficient or the query is out of scope |
| No evaluation shown | Golden dataset and a 7-metric deploy gate |
| Vector search only | Hybrid BM25 + dense + RRF + cross-encoder rerank |
| No audit | Hash-chained, tamper-evident audit ledger for every query |
Deploy Gate
Every metric is a hard block: groundedness ≥ 0.98 · false-confident rate = 0 · citation validity = 100% · recall@10 ≥ 0.90 · correctness ≥ 0.85 · abstention precision ≥ 0.95 · hallucination ≤ 2%.
The groundedness judge is deterministic and independent of whichever LLM does synthesis, so the trust signal cannot collude with the generator.
An independent educational project over public and sample documents — not affiliated with any financial institution, and not investment advice.
Tech Stack
Python · FastAPI · Streamlit · BM25 · Dense retrieval · Cross-encoder reranking · Gemini / Claude / OpenAI adapters · SQLite · Docker · Railway