Amazon Bedrock adds single-shot RAG and retrieval-only via AgenticRetrieveStream

The update addresses a practical gap in retrieval-augmented generation: not every question deserves the same amount of search effort. A direct checklist lookup and a cross-policy comparison can now be served by the same assistant and the same API, with the application choosing the iteration limit per request. The post uses a fictional banking policy assistant to illustrate the three workflows—single-shot RAG for well-scoped questions, agentic retrieval for multi-part comparisons, and retrieval-only when an employee wants to review source text. Iterations are counted as foundation model calls, with the final call reserved for answer generation, so a value of n allows up to n minus 1 planning rounds. Speculative retrieval ahead of the first planning step does not count against the limit. The post advises teams to evaluate the lowest iteration limit that meets each workflow's quality requirements against representative questions, rather than assuming a higher value always produces a better answer.