Financial RAG Chatbot

Analysts and investors spend hours sifting through SEC filings — 10-Ks, 10-Qs, 8-Ks — to extract insight about performance, risk factors, and management commentary. The documents are dense, often 100+ pages, and written in complex legal-financial language.

The goal: an assistant that answers natural-language questions about company financials with source-grounded responses — eliminating the hallucinations that make standard LLM answers worthless in a financial context.

• SEC filings have complex nested structure — tables, footnotes, cross-references — that naive chunking destroys
• Financial data demands exact numbers; approximations are unacceptable
• Relevant information spans multiple document sections, but context windows are finite
• The system must handle both quantitative queries ("What was Q3 revenue?") and qualitative ones ("What are the main risk factors?")
• Every response must cite its sources, or it can't be trusted or audited

A RAG architecture where the retrieval layer does the heavy lifting:

1. Structure-aware ingestion — SEC-EDGAR APIs fetch filings; chunking respects document structure, preserving tables and section boundaries instead of splitting mid-table
2. Semantic search — ChromaDB vector store with text-embedding-3-large, combined with metadata filtering on company, filing date, and section type
3. Multi-stage retrieval — broad initial retrieval, then reranking to surface the most relevant chunks per query
4. Constrained generation — the generator may only answer from retrieved context, and must cite; automatic ticker and period parsing routes queries to the right filings

• FastAPI backend — document ingestion, query processing, response generation
• ChromaDB — persistent vector store with company/filing metadata for filtered retrieval
• LangChain orchestration — the RAG pipeline, conversation memory, chain-of-thought reasoning
• Streamlit frontend — chat interface with conversation history and source highlighting
• Cloud Run deployment — live, with persistent conversation history

Trust in a RAG system is an eval problem, so evaluation is part of the build, not a postscript:

• A multi-model evaluation pipeline uses Claude Opus as judge, scoring responses on accuracy, relevance, and faithfulness to the source documents
• Answers carry line-level citations back into the filing, so any number can be checked against the original in one click
• The corpus spans multiple years of filings across companies, exercising retrieval across heterogeneous document structures

• Time-to-insight cut from hours of manual reading to seconds of conversation
• Automatic ticker and period parsing makes queries frictionless
• Deployed live on Cloud Run with persistent conversation history