Financial RAG Agent

Week 1 implements the P0 retrieval baseline from FINANCIAL_RAG_AGENT_GLOBAL_WORKFLOW.md.

Current scope:

• fixed local sample corpus
• basic metadata ingestion
• BM25 lexical retrieval
• retrieval recall / precision / gold page hit
• wrong company / wrong year checks
• failure report

Run:

python -m src.ingestion.run --config configs/ingestion/sample.yaml
python -m src.retrieval.index --config configs/retrieval/bm25_sample.yaml
python -m src.evaluation.run --config configs/evaluation/sample_week1.yaml
python -m src.reports.build --run-id latest

Week 1 intentionally does not implement answer generation, citation verification, vector retrieval, reranking, XBRL, valuation, or permissions.

Release notes:

• v1.0.0 records the completed Week 1 retrieval baseline.
• Latest validated run: 20260505T154304Z.
• The release validates retrieval only; answer quality, citation faithfulness, and hallucination rate are later phases.
• Week 2 plan defines vector retrieval, hybrid retrieval, reranking, and comparison metrics.

Real FinanceBench / SEC Week 1 run:

python -m src.data_acquisition.download_real --output-dir data/raw/real --sec-tickers MMM,AMD,BBY,AMCR,PEP,BA,PFE,ADBE,VZ,GLW,CVS,GIS,MGM,NKE,AES,AMZN,AWK,XYZ,KO,JNJ,LMT,WMT,MSFT,NFLX,ULTA,COST,KHC
python -m src.ingestion.financebench --config configs/ingestion/financebench_real.yaml
python -m src.retrieval.index --config configs/retrieval/bm25_financebench_real.yaml
python -m src.evaluation.run --config configs/evaluation/financebench_real_week1.yaml
python -m src.reports.build --run-id latest

The real Week 1 corpus uses FinanceBench evidence pages as retrieval chunks and downloads SEC company ticker / submissions metadata for the selected public companies.

Static progress UI:

Open ui/index.html in a browser to view the v1.0.0 Week 1 progress snapshot. The UI has no external dependencies.