AWS walkthrough: agentic retrieval splits multi-intent queries into sub-searches on Bedrock
AWS ML Blog · rss · 2026-10-05
AWS ML Blog demonstrates agentic retrieval on Amazon Bedrock Managed Knowledge Base with LangChain. The core problem: standard RAG packs all user intents (e.g., comparing two products across three dimensions = six sub-questions) into a single query vector, so retrieved chunks look relevant but cover only part of the ask.
Agentic retrieval plans the retrieval instead: it decomposes the question into sub-queries, runs them, judges whether evidence suffices, and searches again if not. The post compares the Retrieve API (single hybrid search) against the AgenticRetrieveStream API (multi-step planning loop streaming trace events), covering cost differences between the two paths and when the cheaper one is right.
Practical details: requires boto3 ≥1.43.32 and langchain-aws ≥1.6.3; IAM must separate the knowledge base service role from the caller identity, noting bedrock:AgenticRetrieveStream and bedrock:InvokeModelWithResponseStream cannot be scoped to a knowledge base ARN; the often-missed bedrock:GetDocumentContent permission is required when the planner's FullDocumentExpansion step pulls a whole document, or queries fail midway.
More from coding & agent
- Teknium catalogs 305 open, hackable devices your AI agent can control and live in — Teknium · 2026-10-06
- Crawler + GPT agent scanned 5,137 jobs for $6.53, found 3 low-competition niches — Arindam_1729 · 2026-10-06
- xAI Cookbook adds 4 Grok SDK examples: screenshot-to-React, X sentiment tracker, AI ad generator — tetsuoai · 2026-10-06
- Power user runs a swarm of personal agents — poke, instinct, muse, dot, Grok bot, pickle and more — sebkrier · 2026-10-06
- omni-macos: A Fully Local Semantic Finder for Apple Silicon, Now Daily-Driver Good — JinaAI_ · 2026-10-06
- User: Team Grok Bots let you spin up org-wide custom tools and workflows in minutes — aryamankhawow · 2026-10-06