Exploring Next
Exploring Next — Ep 332 w/ Justy & Cody — RAG precision tuning can quietly cut retrieval accuracy by 40%, putting agentic pipelines at risk
Justy and Cody dig into new Redis research showing that fine-tuning RAG embeddings for sentence-level precision can quietly hurt general retrieval, sometimes by a lot. They unpack why that matters more in agent pipelines, where one bad retrieval can snowball into bad downstream actions, and why common fixes like hybrid search, MaxSim reranking, or bigger models don't really solve the structural problem. The episode lands on a practical takeaway: keep recall fast, add a separate verification step when correctness actually matters.