Bedrock Knowledge Bases is AWS's managed retrieval-augmented generation (RAG) service: it searches your proprietary data for information relevant to a query and generates answers with citations back to the original sources. Claude is explicitly supported as the generation model — AWS's own API examples pass Claude model ARNs and inference profiles to the RetrieveAndGenerate API. But before you pick a generation model, you face an architectural fork the AWS docs now draw sharply: managed versus customer-managed knowledge bases.
The two architectures side by side
| Managed Knowledge Base | Customer-managed | |
|---|---|---|
| Who runs ingestion, indexing, storage, retrieval | Bedrock, end to end | You configure the pipeline |
| Vector store | Service-managed (invisible to you) | Yours: OpenSearch Serverless, Aurora, Neptune, and others |
| Embedding / re-ranking models | Service-managed by default | You choose (Amazon Titan or Cohere Embed families) |
| Retrieval style | Agentic multi-hop retrieval | Classic retrieve, or retrieve-and-generate |
| Query APIs | AgenticRetrieveStream or Retrieve | Retrieve or RetrieveAndGenerate |
| MCP integration | AgentCore Gateway | Not documented |
AWS recommends the managed type. With it, Bedrock handles ingestion, indexing, storage, and retrieval, using service-managed embedding, re-ranking, and reasoning models by default, and adds agentic multi-hop retrieval — the system can chain multiple retrieval steps to answer a question no single lookup covers. Managed knowledge bases connect to Amazon S3, SharePoint, Confluence, Google Drive, OneDrive, and a Web Crawler, and they enforce document-level ACL permission filtering at retrieval time (for every connector except the Web Crawler) — a significant governance feature if different employees are entitled to see different documents.
What managed buys you: less pipeline, plus an MCP door
The operational pitch is straightforward: no vector database to size, patch, or pay for while idle, no chunking-and-embedding pipeline to maintain, and retrieval quality improvements arrive as service updates rather than as your backlog. The distinctive extra is the AgentCore Gateway MCP integration: managed knowledge bases can be exposed through AgentCore Gateway using the Model Context Protocol (MCP), the open standard that lets agents discover and call tools. That turns your knowledge base into a tool an agent can query — including agents built on Claude — rather than a service only your application code calls.
The trade is control and visibility. You do not pick the embedding model, you cannot inspect or directly query the vector index, and you cannot point existing search infrastructure at it. Note one API asymmetry that surprises teams migrating between the two types: RetrieveAndGenerate — the one-call "search and answer with citations" API — cannot be used with managed knowledge bases. Managed knowledge bases are queried through AgenticRetrieveStream or plain Retrieve instead.
What customer-managed buys you: your store, your rules
With a customer-managed knowledge base you control the pipeline and the vector store — OpenSearch Serverless, Aurora, or Neptune, among others. That matters when you already operate one of those stores, when your security team wants the index inside infrastructure they can audit directly, or when you need explicit control over chunking strategy and embedding model choice (the documented options are Amazon Titan and Cohere Embed families; Voyage embeddings, which Anthropic's docs point to, are not in the Knowledge Bases list). You also keep RetrieveAndGenerate, whose responses map each generated text span to the retrieved passages that support it — see the RetrieveAndGenerate deep dive — with session continuity, guardrails, prompt templates, metadata filters, and reranking configuration.
Where Claude sits in either design
In both architectures, Claude serves as the generation (and optionally parsing) model: Claude vision models are explicitly listed as usable parsers for advanced document parsing, and Claude works as the response generator via model ARNs or inference profiles (which enable cross-Region inference — with the documented caveat that cross-Region inference can move data across Regions). One framing note: Knowledge Bases runs on Bedrock's own agent-runtime APIs with anthropic.-prefixed model IDs, not the Anthropic Messages API surface. Since Anthropic's server-side web search and web fetch tools are unavailable on Bedrock, Knowledge Bases is the Bedrock-native grounding path.
Where to go next
See Knowledge Base data sources for connector details, multimodal chunking for audio and video ingestion, and RAG basics if you want the concepts from first principles.