What a Single Query Vector Averages Away
A question with several intents gets one search. Similarity search compresses the whole request into a single query vector, and the retriever returns the best approximation of the average of those intents (S14E3). The chunks that come back can be topically relevant and still cover only a fraction of what the question actually asked (S14E5). AWS describes agentic retrieval on Amazon Bedrock Managed Knowledge Base as replacing that single search with a plan: break the question into sub-queries, run them, judge whether there is enough evidence, and search again if there isn't (S14E10). More searches are the price of coverage, which makes the practical question not whether agentic retrieval is better but when the extra work pays for itself.
AI-generated discussion · Different perspectives
Averaging isn't automatically wrong. If the parts of a question pull in the same direction, collapsing them into one vector is a bargain — one search, one pass, done.
The bargain is the problem when the parts disagree. The answer stays smooth while the evidence behind it thins out, and nothing in the interface signals the difference.