FinSight
Question answering over financial reports, with page-level sources
Answers over financial PDFs with the source page attached to each one
The problem
Annual reports and quarterly filings run to hundreds of pages. Analysts need the figure, the sentence it came from and the page number, not a summary they cannot verify.
What I built
- Ingestion of report PDFs with PyPDF or pdfplumber for table-heavy pages, split into 1,000-character chunks with 200 characters of overlap
- Embeddings with text-embedding-3-small stored in a persistent ChromaDB collection with page metadata
- Top-five retrieval into GPT-4 at low temperature, answers returned with the source documents, page numbers and a preview of the passage used
- A CLI with an interactive mode, single-question mode, and preset revenue and strategy analyses
Results
Evaluation pending. A twenty-question comparison of this retriever against a hybrid retriever (BM25 plus embeddings) with a cross-encoder reranker is being added. Until that table exists, no accuracy claim is made for this project.
Limitations
- No published evaluation yet
- Single embedding retriever, no reranking
- Command-line interface only
What this means for your project
This is the baseline shape of the document assistant service. A pilot adds your documents, twenty of your questions and the scored table that this project is still missing.