Agentic retrieval with LangChain and Amazon Bedrock Knowledge Bases

This article demonstrates agentic retrieval with LangChain and Amazon Bedrock Knowledge Bases, showing how it improves upon standard retrieval for complex, multi-part questions.

• Standard retrieval uses a single query vector that misses multiple intents; agentic retrieval decomposes questions into sub-queries
• Amazon Bedrock Managed Knowledge Base handles chunking, embedding, storage, and retrieval without self-managed vector stores
• Two APIs available: Retrieve (single-shot, cheaper, faster) and AgenticRetrieveStream (multi-step planning with trace events)
• Agentic retrieval plans queries, evaluates evidence sufficiency, and re-retrieves if needed across up to five knowledge bases
• Trace events show the planning loop including SpeculativeRetrieval, Planning, Retrieval, and FullDocumentExpansion steps
• Use standard Retrieve for short, well-scoped questions; use agentic retrieval for multi-part, comparative, or exploratory questions
• Includes code examples for creating knowledge bases, querying with LangChain, and reading trace events via boto3

Agentic retrieval trades higher per-call cost for improved recall on complex questions by using built-in model-based query planning.