Amazon Web Services has made Amazon Bedrock Managed Knowledge Base generally available. The service is designed to give AI agents and generative AI applications access to enterprise information without requiring teams to assemble and operate separate connectors, parsers, vector stores, retrieval models and access-control layers.
AWS says customers can create a knowledge base, connect a source and begin ingestion from the management console without selecting a model. The service applies defaults for embedding, storage and chunking, while retaining configuration options for organisations that want to choose their own embedding model, reranker or chunking strategy.
Six native connectors and direct ingestion
At general availability, Managed Knowledge Base includes native connectors for Amazon S3, Microsoft SharePoint, Atlassian Confluence, Google Drive, Microsoft OneDrive and web crawling. A direct ingestion API covers documents stored elsewhere. Subsequent synchronisations process changed or new documents rather than rebuilding the entire collection.
The service supports access-control lists and performs real-time permission checks against the authoritative source in addition to pre-retrieval filtering. AWS says the documents examined during that check are transient for the API call and are not exposed to the model or end user. Organisations should still test permission mappings and failure modes against their own identity systems before relying on the control in production.
Managed parsing and storage
Managed Knowledge Base can parse standard digital documents as well as tables, charts, diagrams, scanned material, audio and video. AWS lists support for visual documents such as PDF, PowerPoint and Word files up to 500 MB, audio files up to 2 GB and video files up to 10 GB.
After parsing, the service segments content for retrieval. It can select a chunking strategy automatically, or customers can specify fixed-size chunks or disable chunking for material that has already been prepared. The storage layer is also managed: AWS provisions and scales the underlying retrieval storage, keeps hybrid keyword and semantic search enabled and handles operational work such as monitoring, backup, patching and capacity management.
Direct and agentic retrieval
The service exposes two retrieval patterns. The Retrieve API returns ranked source chunks for direct questions and latency-sensitive uses. Agentic Retrieval uses a foundation model to split a complex question into sub-queries, search one or more knowledge bases, evaluate whether the evidence is sufficient and repeat the process when needed.
Agentic Retrieval can run up to five iterations by default, with a configurable maximum. It streams trace events that show planning and retrieval steps before returning deduplicated results. Customers choose the model used for planning and evaluation, which means accuracy, latency and cost will vary with both the model and iteration settings.
Integration with AgentCore Gateway
Managed Knowledge Base integrates with Amazon Bedrock AgentCore Gateway. A knowledge base can be presented as a tool that Model Context Protocol-compatible agents discover through the gateway, while AWS Identity and Access Management, routing and observability are centralised. The service can also be called directly without AgentCore.
This integration is intended to reduce coupling between agent code and individual knowledge-base identifiers. Frameworks including Strands, LangChain and CrewAI can use the standardised endpoint, subject to the organisation's gateway configuration and IAM policies.
What buyers should verify
General availability removes several infrastructure decisions, but it does not remove the need to evaluate retrieval quality, permissions, source freshness and model behaviour. Teams should test difficult multi-hop questions, documents with complex layouts and cases where a user's access changes between ingestion and retrieval.
The announcement does not provide a complete regional availability and pricing table. Customers should confirm supported regions, connector limits, ingestion and storage charges, model costs and data-governance requirements in the current AWS documentation before deployment. The main benefit is operational consolidation; its value will depend on whether the managed defaults produce acceptable retrieval quality for an organisation's actual corpus.