Skip to content

AWS adds Bedrock knowledge base migration skill for coding agents

A new guide in the amazon-bedrock skill lets a coding agent plan and stage a move from a customer-managed knowledge base to a managed one.

Draftship Editorial3 min read

A migration guide that ships inside the skill

On 12 November 2026, AWS published a walkthrough for moving a customer-managed knowledge base onto the Amazon Bedrock Managed Knowledge Base with a coding agent doing part of the work. The mechanism is a migration guide bundled into the amazon-bedrock skill, which ships with the Agent Toolkit for AWS. The point is to give the agent enough context to plan the move with you instead of improvising it.

The flow starts read-only. The agent discovers the existing configuration and produces a plan you review line by line. Only after approval does it act, either creating the new knowledge base and its data sources alongside the original, or writing a migrated infrastructure-as-code template for your pipeline. Retrieval is validated before any traffic is switched.

What the managed option takes over

On a customer-managed knowledge base, you provision and scale the vector store, tune the embedding and index settings, and plan capacity for each data source you add. Managed Knowledge Base hands ingestion, storage, indexing and retrieval to Bedrock and scales the store as content grows. Hybrid search, which combines keyword and semantic matching, stays on with no index to configure. A managed embedding model, a managed reranker and a parser for multimodal file types are part of the service.

You can still bring your own Amazon Bedrock embedding model at creation time, and each retrieval request can use the managed reranker, an Amazon Bedrock reranking model, or no reranking. The managed reranker requires the managed embedding model.

According to the post, three capabilities exist only on the managed side: agentic retrieval, which handles multi-part questions iteratively and can search more than one knowledge base in a single request; exposure as a target through AgentCore Gateway so MCP clients can query it with the gateway handling authentication; and resource-based policies for cross-account access, scoped to specific roles and actions such as bedrock:Retrieve.

Connectors cover sources including Amazon S3, Microsoft SharePoint and Atlassian Confluence. After the first full sync, later syncs take only changes, on demand or on a schedule, and a deletion safeguard protects against accidental bulk removal of indexed content. ACL awareness carries document permissions into retrieval, with a real-time check against the source for connectors that support it. It does not authenticate users, so the application has to verify and pass the user's identity.

Quotas and what you pay for

Default quotas change noticeably. Knowledge bases per account per Region go from 100 to 10,000, data sources per knowledge base from 5 to 200, concurrent ingestion jobs per knowledge base from 1 to 50, and query input characters from 1,000 to 10,000.

On cost, the post puts managed vector storage in a base storage rate of $5 per GB per month, with retrieval metered, so spend follows usage. A customer-managed setup is billed by whichever vector store you picked, and some charge for provisioned capacity even when idle. Amazon S3 Vectors is cited as an example that charges only for what is used. Multimodal parsing on a customer-managed knowledge base needs Amazon Bedrock Data Automation, billed by page or image.

Limits and open questions

The agent does not finish the job. Ingesting content and updating the application's Retrieve calls stay with your team, although the agent can walk you through both. Prerequisites and supported Regions live in linked documentation rather than in the guide itself, and the post does not name which coding agents were tested. It is also unclear whether the higher default quotas apply to existing accounts immediately, and what retrieval metering adds beyond the storage rate.

What this means for building apps with AI

Most of the effort here is infrastructure, not application code: the retrieval call in your app changes, but the interface around it does not. Teams weighing the move should treat the agent's plan as a review artifact rather than an automatic switch, since ingestion and cutover remain manual steps. Draftship is not affiliated with AWS, and nothing here suggests it supports Bedrock knowledge bases. If you are wiring model calls into an app of your own, the documentation on adding AI to your app covers that separate piece of the problem.

Primary sources

This article was drafted with AI assistance from the primary sources listed below and checked automatically against them. An editor hasn't reviewed it yet.