Lucy hits 94.7% Recall@10 on FinanceBench with Qdrant over 2.8M SEC/DART filings
qdrant_engine · x · 2026-10-03
Financial AI company Lucy built a source-of-truth retrieval layer over 2.8M+ SEC EDGAR and Korea DART filings on Qdrant.
- Four hybrid retrieval lanes, with payload filters applied inside the query rather than after retrieval
- 94.7% Recall@10 on FinanceBench
- Retrieval improvements alone added +0.160 answer accuracy with the answer model held fixed
- An open-weight model on Lucy RAG beat frontier models doing web search
A Qdrant engineering case study in building reliable RAG for high-stakes financial data.
More from coding & agent
- Runway MCP one-shots a corporate training video in 15 minutes, typically $10k to make — tlakomy · 2026-10-03
- Building an AI Research System for a One-Person Firm: Progressive Determinism Architecture — otionflow · 2026-10-03
- Dev deletes custom CLI after discovering SmartSub v3.9.0 already shipped CLI + MCP — lxfater · 2026-10-03
- Meta reportedly building Agent Engine codenamed Forge for its API platform — testingcatalog · 2026-10-03
- Architecture for a One-Person Research Business: Progressive Determinism Around Codex — DutyOnly4308 · 2026-10-03
- Where should the agent end and the harness begin? A Rust runtime builder's open question — sreevarshan-xenoz · 2026-10-03