FinRAG — Finance-Domain RAG System
A retrieval-augmented generation system for querying SEC filings and earnings transcripts with grounded, source-cited answers — no hallucinated numbers.
How It Works
Financial documents like 10-Ks and earnings transcripts are long, dense, and full of domain-specific language that generic RAG systems handle poorly. FinRAG combines a finance-aware ingestion pipeline with a retrieval pipeline tuned for precision over recall, so answers stay grounded in the source document.
Ingest
PyMuPDF + pdfplumber fallback
Chunk
Finance-aware, table-preserving
Embed
OpenAI text-embedding-3-small
Store
ChromaDB, SHA256-deduped
Rerank
MMR, built from scratch
Answer
GPT-4o-mini, source-cited
Try It Live
Ask a question about the indexed financial filings — or upload your own document — and get a grounded, source-cited answer from the live pipeline below.
Hosted on Render's free tier — if the server has been idle, the first query can take 20–30s to wake up. Subsequent queries are fast.
Answer
Your document is processed in memory only and automatically discarded after about 30 minutes — it is never stored permanently or added to the shared corpus. This is a public demo with no login: please do not upload documents containing SSNs, full account/card numbers, or other highly sensitive personal information.
Answer
Powered by FinRAG — OpenAI embeddings · ChromaDB · MMR reranking · GPT-4o-mini. Answers are grounded in the indexed filings and may not reflect all available data.