Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseInsight · On RadarAWS Machine Learning Blog→ original

AWS Teaches a Coding Agent to Move Bedrock Knowledge Bases to a Managed Version

On Amazon Bedrock, teams had to run their knowledge bases themselves: choose a vector store, configure embeddings, and estimate capacity. AWS has added a guide to Agent Toolkit that lets a coding agent study your knowledge base, write a plan, and create a managed copy next to it without touching the original. Storage costs $5 per GB per month, and search costs $1 per 1,000 requests. For rarely used knowledge bases, this may be cheaper, but starting indexing and switching traffic remains the team's job.

AI-processed from AWS Machine Learning Blog; edited by Hamidun News
AWS Teaches a Coding Agent to Move Bedrock Knowledge Bases to a Managed Version
Source: AWS Machine Learning Blog. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceAWS Machine Learning Blog ↗
Editor-in-Chief
Key metric #1The new MANAGED-type knowledge base is created alongside the original, which the agent does not change or delete
Limits are higherup to 200 data sources instead of 5, and 50 parallel ingestion jobs instead of one
Key metric #3Agentic search costs another $4 per 1,000 requests, for a total of roughly $6–9

If you build RAG on Amazon Bedrock, you know what it costs to run your own knowledge base. You have to choose and deploy a vector store, pick an embedding model, configure the index, and estimate capacity for each new data source. AWS is offering to take that work off the team's hands and move to Managed Knowledge Base. To keep the move from turning into weeks of manual work, the company has added a migration guide to the amazon-bedrock skill in Agent Toolkit for AWS. It is a set of instructions that your coding agent (for example, Kiro) uses to plan and partly carry out the migration.

What changes in the move

In the managed version of Bedrock, the service itself handles ingestion, storage, indexing, and search. Hybrid search, which combines keyword matching and semantic similarity, is always on, and you don't need to configure the index. There is a built-in embedding model, a built-in reranker (which re-sorts retrieved chunks by relevance), and a parser for multimodal files. You can still connect your own embedding model from Bedrock, but then the managed reranker will not work.

There are also features unavailable in knowledge bases you manage yourself. Agentic search breaks down compound questions and searches step by step, including across several knowledge bases at once. The AgentCore Gateway integration exposes the knowledge base to any client that supports MCP. Resource policies let applications in other AWS accounts use someone else's knowledge base. Native connectors pull data from S3, SharePoint, and Confluence, and during sync they pick up only changes. Access rights are also respected: search returns only what the signed-in user is allowed to see. However, the service does not handle user authentication; your application has to pass the identity.

Default quotas have grown considerably. You can create 10,000 knowledge bases per account and region instead of 100, 200 data sources per knowledge base instead of 5, 50 parallel ingestion jobs instead of one, and a query can contain 10,000 characters instead of 1,000. This means each team or product can get its own knowledge base.

What it costs

Pricing in the managed version is tied to actual usage. Storage, including vectors, parsing of multimodal documents, default embeddings, syncs, and connectors, is covered by the base rate of $5 per GB per month. Standard search costs $1 per 1,000 requests, and reranking is already included. Agentic search adds another $4 per 1,000 requests, so with the base calls the total comes to roughly $6–9 per 1,000. According to AWS's estimate, building the same agentic workflow yourself on a Bedrock model costs $18–24 per 1,000 requests, before development and maintenance costs.

In knowledge bases you manage yourself, some vector stores charge for dedicated or minimum capacity even without traffic. As a result, internal knowledge bases for support, policies, and regulations that are used infrequently are cheaper to model on a pay-per-use basis. The final conclusion depends on your volumes, and AWS suggests the agent run a cost analysis using CloudWatch metrics, S3 object sizes, and Cost Explorer data. Each figure in the analysis is labeled as measured, provided by you, or estimated.

How the migration works

The guide separates planning from execution. In Explore mode, the agent answers questions and reads configuration in read-only mode so you can assess the effort. In Migration mode, it goes through checkpoints: it explains what it will and won't touch, confirms the account and region, examines the knowledge base or IaC templates (CloudFormation, CDK, Terraform), and drafts a plan. By default, it stops there.

In the plan, each setting is marked as one-to-one portable, replaced, or blocked. For example, the vector store will become automatic, semantic chunking will be replaced by the managed default, and S3 prefixes and metadata will carry over unchanged. If there is no migration path, the agent excludes that knowledge base from the build and advises contacting AWS Support.

If you approve resource creation, the agent proposes an IAM policy with an explicit deny on modifying the source knowledge base, shows the script, and runs it only if the plan hash matches. The new MANAGED-type knowledge base appears alongside the original, which continues to serve traffic. The script does not start ingestion, does not call Retrieve, and deletes nothing. The agent is also meant to ignore instructions found in readable data, which protects it against simple prompt injection. All calls can be audited in CloudTrail.

What remains your job

Loading content, changing the Retrieve call in the application, and switching traffic remain the team's responsibility. In the application, you need to change the knowledge base identifier and replace vectorSearchConfiguration with managedSearchConfiguration. You can no longer force pure semantic search; search is always hybrid. Before switching, it's worth comparing search quality on your own documents, especially because of the change in chunking. For the IaC path, don't forget to run terraform init and validate manually and keep the original resources for rollback. One caveat: some settings, such as enabling ACLs on S3, cannot be changed after the knowledge base is created, so decide on those in advance.

Story

Key Questions Answered

What changes in the move?

In the managed version of Bedrock, the service itself handles ingestion, storage, indexing, and search. Hybrid search, which combines keyword matching and semantic similarity, is always on, and you don't need to configure the index. There is a built-in embedding model, a built-in reranker (which re-sorts retrieved chunks by relevance), and a parser for multimodal files. You can still connect your own embedding model from Bedrock, but then the managed reranker will not work.

What it costs?

Pricing in the managed version is tied to actual usage. Storage, including vectors, parsing of multimodal documents, default embeddings, syncs, and connectors, is covered by the base rate of $5 per GB per month. Standard search costs $1 per 1,000 requests, and reranking is already included. Agentic search adds another $4 per 1,000 requests, so with the base calls the total comes to roughly $6–9 per 1,000.

How the migration works?

The guide separates planning from execution. In Explore mode, the agent answers questions and reads configuration in read-only mode so you can assess the effort. In Migration mode, it goes through checkpoints: it explains what it will and won't touch, confirms the account and region, examines the knowledge base or IaC templates (CloudFormation, CDK, Terraform), and drafts a plan. By default, it stops there. In the plan, each setting is marked as one-to-one portable, replaced, or blocked.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…