sec-filing-agent

Vouch is a finance-grade tool-calling agent that answers financial questions about any U.S. public company using SEC filings. It follows a dual‑path architecture: exact figures come from deterministic tools over live XBRL data (the LLM never performs arithmetic on financials), while narrative context is retrieved via pgvector dense search over 10‑K, 10‑Q, and 8‑K documents. An output guardrail blocks any number that cannot be traced to its source—either an XBRL tool or a cited filing passage—ensuring the system abstains rather than fabricates. Key capabilities include tools for fetching exact figures (`get_financials`), standard ratios (`get_ratio`), year‑over‑year growth (`get_growth`), custom formulas (`compute_formula`), full statements (`get_statement`), largest/smallest line items, segment/geography breakdowns, and qualitative search. Ratios and growth are computed with fixed conventions to avoid the common LLM error of picking the wrong base metric. Delisted or renamed firms are resolved by name, not a dead ticker. The system is validated through a three‑layer evaluation: an internal CI gate of 84 cases scoring numeric grounding, citation, abstention, and trajectory metrics; calibrated LLM judges (faithfulness at Cohen’s κ = 0.76, zero false positives on abstaining answers); and external benchmarking on FinanceBench (Patronus AI), where it achieved 94% addressable coverage on the numeric set at a zero‑fabrication rate and 96% narrative groundedness. Adversarial red‑team testing drove the fabrication rate from 11% to 0. Users can upload their own financial documents (internal reports, non‑public companies) and the agent answers from them using table‑cell extraction via Docling, with cell‑level citations (`filename · page · row/col`). Each document is per‑user isolated and scoped to prevent cross‑file attribution. Deployment exposes a local service with `/ask`, `/export`, `/upload`, and an OpenAI‑compatible chat endpoint. The default model is o4‑mini, and the system includes a miss queue that logs unhandled metrics for future expansion. Vouch is built on LangGraph and pgvector, reusing retrieval and CI‑eval engineering from the earlier Slug Advisor project.