Turning financial filings into a question-answering copilot

Client / Context
Internal project for asset-management teams who live in SEC filings, earnings transcripts, and research PDFs.

Problem
Analysts were wasting hours searching, skimming, and copy-pasting from long documents. Existing search was keyword-based and missed tables, footnotes, and subtle wording.

What we built
We designed and implemented FinDocs Online, a financial-document Q&A system:

• Upload 10-K/10-Q filings, call transcripts, and PDF research notes
• Ask natural-language questions like “What are the main drivers of margin expansion in 2024?”
• Hybrid retrieval (BM25 + dense embeddings) with table-aware chunking to handle numerical data
• Answers always come with source paragraphs and page locations, not just free-form text
• Simple reasoning on top of numbers (growth rates, YoY/ QoQ deltas, simple aggregations)

Outcome
On a small internal test set, FinDocs achieved high hit@K and near-exact match scores, and most importantly, analysts reported they could:

• Find the right paragraph in seconds instead of minutes
• Trust the answer, because every claim has a visible citation
• Use FinDocs as a daily “copilot” during earnings season

‍