Adding a vector database is usually the first move when output quality drops on a knowledge-heavy task. It's rarely the right one. Most quality problems in that category aren't retrieval failures, they're curation failures wearing a retrieval-shaped disguise. The model isn't underinformed. It's drowning in loosely relevant material with nothing telling it what to weigh more heavily, and RAG just automates feeding it more of exactly that.
The tell is what happens after teams add retrieval and the problem doesn't fully go away, just shifts shape. Answers get more grounded in the sense that facts are technically present in context, but they get vaguer in the sense that matters, the model still can't tell which of the five retrieved chunks is actually load-bearing for this specific question. Retrieval expanded the pool of correct information without ever solving the part where the model has to decide what to do with it. That's not a retrieval problem. That's the same context-structuring problem RAG was supposed to fix, just relocated one layer downstream.
Where RAG earns its complexity is genuinely large, frequently changing corpora where you can't fit the relevant slice into context even after aggressive curation, legal discovery, large codebases, that kind of thing. For a lot of internal tools and product features, the actual fix is smaller and less interesting than a vector database: cut the source material down to what's structurally relevant to the task before it ever reaches the model, and be more deliberate about what "relevant" means for that specific request. Teams skip that step because it requires someone to actually think about the data, and reach for retrieval infrastructure instead because it's a known pattern with existing tooling.
Not claiming RAG is never the right call. Claiming it's reached for by default in cases where the actual bottleneck is upstream of retrieval entirely. If someone's shipped a case where RAG measurably fixed a quality problem that better context curation alone couldn't have, genuinely interested in what that looked like.
[link] [comments]