Amazon Bedrock Managed Knowledge Bases: agentic hits 6 of 6

Amazon Bedrock Managed Knowledge Bases removes the self-managed vector store, embeddings, and re-ranking models from the RAG architecture; the walkthrough uses an S3 bucket as the data source, and the API takes no storageConfiguration, which the post calls the clearest signal that Amazon handles the storage layer. From there the same complex query can go through two paths: a standard LangChain retriever wrapping the Retrieve API, or a planning loop exposed as agentic_retrieve and the lower-level agentic_retrieve_stream call. The post notes friction in the current integration — agentic retrieval is a function rather than a LangChain retriever, so it needs a RunnableLambda to sit in a chain, and trace events that show the query plan require a direct boto3 call. For context, AWS evaluated agentic retrieval on MuSiQue, a public multi-hop benchmark, and reported improved recall over single-shot retrieval, with the largest gains on the hardest questions and single-hop questions seeing gains under five points. The post also notes that agentic retrieval registers up to five knowledge bases in one request and routes sub-queries using a natural-language description attached to each, which the other API cannot do. The outcome likely hinges on whether a team's own query mix is dominated by single-intent questions, since the recommendation is to route on query shape rather than pick one path for everything, and measuring that mix appears to be the intended next step.