Agentic retrieval for Amazon Bedrock Managed Knowledge Base

This article introduces agentic retrieval for Amazon Bedrock Managed Knowledge Bases, a new API that handles multi-part, comparative, and exploratory questions by decomposing queries and iterating over retrieval steps.

• Single-shot retrieval fails on multi-intent questions; agentic retrieval decomposes queries and retrieves per intent
• AgenticRetrieveStream API runs a planning loop driven by a foundation model with visible trace events
• Supports routing across up to five knowledge bases with natural-language descriptions for intelligent KB selection
• Benchmarks show 20% absolute recall improvement on multi-hop questions, with larger gains on harder questions
• Priced per call ($4 per 1,000 calls for managed model) plus underlying retrieval costs
• Best practices include small planner models, maxAgentIteration tuning (3 for single-KB, 4-5 for multi-KB), and CloudWatch logging
• Supports optional response generation, reranking, and guardrails (BLOCK action only)

Agentic retrieval moves query planning and iteration into the Bedrock service, eliminating custom orchestration code while providing consistent security, observability, and cost behavior.