RAG solves a problem most teams don't have: the real bottleneck is context curation

ClickOk5811 · reddit · 2026-08-22

A high-signal Reddit post argues that when output quality drops on knowledge-heavy tasks, teams reach for a vector database first—and it's usually the wrong move.

The core claim: most such quality problems aren't retrieval failures but curation failures—the model isn't underinformed, it's drowning in loosely relevant material with no signal about what to weigh more heavily. The tell: after adding retrieval, the problem doesn't go away, it shifts shape—facts are technically present in context, but the model still can't tell which of five retrieved chunks is load-bearing for the specific question. That's the same context-structuring problem RAG was supposed to fix, relocated one layer downstream.

Where RAG earns its complexity: genuinely large, fast-changing corpora where the relevant slice can't fit in context even after aggressive curation—legal discovery, large codebases. For most internal tools, the real fix is smaller: cut source material down to what's structurally relevant before it reaches the model. Teams skip that because it requires actually thinking about the data. The author invites counterexamples where RAG measurably fixed a problem that better curation alone couldn't.

Original post →

More from coding & agent

coding & agent channel →