Filings RAG QA Demo

query  ·  ingest

Cited RAG over company filings. Ask about an issuer's quarterly disclosures, get a source-grounded answer with every claim traceable to a specific chunk. Generation is switchable between OpenAI and Anthropic. Plug-and-play: point the ingest notebook at any US filer and it pulls the 10-Q from SEC EDGAR, builds the index, and the query notebook answers over it.

Ingest once, query many times. Three issuers come pre-indexed so the query notebook runs out of the box:

Colab: click a badge above. Local:

git clone https://github.com/FranQuant/filings-qa-rag-demo
cd filings-qa-rag-demo
pip install openai anthropic pyarrow pandas python-dotenv beautifulsoup4 requests
printf 'OPENAI_API_KEY=sk-...\nANTHROPIC_API_KEY=sk-ant-...\n' > .env   # Anthropic optional

Then open either notebook and run all cells. OpenAI key required (embeddings); Anthropic optional (provider comparison); ingestion also needs your email for EDGAR's User-Agent.

import sys; sys.path.insert(0, "src")
from genai_filings.answering import synthesize_answer

answer = synthesize_answer(
    query="How is net interest margin evolving?",
    issuer="JPM", period="2026Q1", k=5,
    provider="anthropic",      # or "openai"
    model="claude-opus-4-8",   # omit for the provider default
    temperature=0.0, max_tokens=500,
)
print(answer["citations_valid"], answer["citations_invalid"])
print(answer["answer_markdown"])

provider selects the generator only (embeddings always use OpenAI). Defaults gpt-5.5 / claude-opus-4-8; newest models drop temperature (handled automatically) — for reproducible runs use gpt-4.1 or claude-sonnet-4-6.

Research demo, not investment advice. Answers come strictly from the retrieved excerpts. EDGAR carries the formal 10-Q/10-K, not call transcripts, so EDGAR issuers lack management Q&A color. Verify table figures against the original filing.