Raggar

Upload a long financial report and ask questions about it in plain English. Every answer cites the passage it came from.

Retrieval

Dense retrieval alone kept missing exact financial terms, because an embedding of "EBITDA" sits near a lot of things that are not EBITDA. Keyword search alone missed anything paraphrased. Raggar runs BM25 and FAISS over all-MiniLM-L6-v2 embeddings together, then reranks the merged candidates with a weighted score before passing the top passages to Llama 3 70B.

The frontend is FastAPI behind React and Vite, with PDF parsing through PyMuPDF, live document viewing, and a chat interface next to the source.

Built with a teammate. The repository is private, so there is no public link for this one.