jev-for-finance

Jev applied to financial research. Jev RAG vs a Vals-style agentic RAG baseline on 10-Ks from Apple, Microsoft, Nvidia and Amazon: 50/50 right with one LLM call and ~100x fewer tokens than the baseline. Code, logs.

Jev for Finance applies Jev to financial research, starting with RAG over SEC filings. Retrieval is treated as a decision: code finds the filing, search keeps 30 passages, Jev picks the one that answers, and an LLM reads two. This repository holds that pipeline (Jev RAG), a benchmark against an agentic RAG baseline built on the Vals AI Finance Agent harness, 50 analyst questions about four 10-Ks, and every log behind the numbers. Full write-up: Is Jev efficient for RAG?