Agentic RAG over Filings & News
Grounding answers in EDGAR filings and news, with chunking and citation discipline doing the heavy lifting.
Retrieval-augmented generation grounds a model in documents instead of memory: chunk filings and news, embed the chunks, retrieve the relevant ones, and cite them in the answer. Agentic RAG adds a loop in which the agent reformulates queries, pulls from EDGAR full-text search, and verifies each claim against the retrieved text before writing. Chunking choices matter more than models: split by section so a paragraph keeps its context, and keep metadata like accession number and filing date attached. Citation discipline is the whole game in finance, because every number in an answer must trace to a source, and an honest agent says it cannot find support rather than guessing.