FinanceRAG

FinanceRAG is a retrieval-augmented generation app you built to answer questions about live financial news rather than relying on a model's frozen training data. It works by embedding articles from the Yahoo Finance Wire and storing them in pgvector, then for each query it runs a nearest-neighbor search that ranks results by a blend of similarity and recency so a highly relevant but outdated story loses out to a fresher one that's still a good match. The top three retrieved passages become the model's entire context, and every answer is composed strictly from those passages, each one dated and linked back to its source. The core problem this solves is the classic weakness of LLMs applied to finance specifically: they don't know what happened today, and they give no way to verify a claim. By grounding every answer in live, timestamped, citable sources, the app trades raw model "knowledge" for traceability and currency both of which matter a lot in a domain where a stale or unverifiable answer can be actively misleading. As for what you personally learned building it, I can only see the finished architecture from the live site, not your process — if you share your code, README, or build notes, I can help you turn the specific technical decisions (like the recency-weighting scheme or pgvector setup) into a sharper "what I learned" narrative.