Agentic RAG: Should the Agent Reason About Where to Retrieve, Not How?

Arc_bong · reddit · 2026-09-04

A Reddit discussion on where agentic RAG actually adds value. The author argues that repeated "search → inspect → search again" loops only make sense for genuine multi-hop questions; with multiple knowledge bases, the agent should reason about where to retrieve (query → route to relevant KBs → parallel retrieval → rerank/merge → answer) rather than micromanaging retrieval itself.

A subtle failure mode: each KB may return good candidates, but the cross-KB fusion layer can still pick wrong context — RRF's rank assumptions break down as sources multiply, and raw similarity thresholds aren't comparable across corpora.

The real design question: what should the agent control, and what stays deterministic? The author compares LangGraph/LangChain, LlamaIndex, and Lyzr's Agent Studio, and asks how production teams draw this boundary.

Related event: Agentic RAG Debate: Route the Query, Don't Just Retrieve(2 posts)→

Original post →

More from coding & agent

coding & agent channel →