AWS compares standard and agentic retrieval with LangChain and Amazon Bedrock Knowledge Bases

Published on October 5, 2026, AWS’s walkthrough builds a Retrieval Augmented Generation (RAG) application with LangChain and Amazon Bedrock Managed Knowledge Base. Using Amazon S3 as the data source, it runs the same multi-part question through standard and agentic retrieval so readers can compare their outputs, trace events, and costs.

The guide presents the paths as serving different query needs. The Retrieve API runs one hybrid search and returns scored chunks. AgenticRetrieveStream runs a planning loop: it splits a question into sub-queries, retrieves for them, checks whether the evidence is sufficient, and streams the steps as trace events. When response generation is enabled, the agentic helper can also return a grounded answer and citations alongside retrieved chunks; its result deduplicates chunks gathered across iterations.

AWS also details the LangChain integration. The langchain-aws package exposes both retrieval modes: standard retrieval is used as a LangChain retriever, while agentic retrieval is a standalone function. The post says agentic retrieval works only with Amazon Bedrock Managed Knowledge Base. Unlike Retrieve responses, its results do not include an equivalent typed relevance-score field. Both paths support guardrails, but agentic retrieval uses a different configuration and supports BLOCK mode only.

For short, well-scoped questions, AWS recommends Retrieve, describing it as faster and less expensive and noting that it also works with self-managed knowledge bases. Agentic retrieval is the alternative for questions that benefit from iterative planning and retrieval. The walkthrough builds on AWS’s July 23, 2026 guide to agentic retrieval, adding a practical comparison centered on LangChain integration and the two retrieval paths.