The complete guide
Amazon Bedrock integration: the complete guide
Updated November 18, 2026 · Published by Danny Rathod, Founder, StackLift AI. Written with AI assistance; how we make content.
Short answer
Amazon Bedrock puts models from several providers behind AWS endpoints, so access is governed by IAM and logged by your AWS tooling. An integration picks the right API (Converse for most chat work), grants a role only the models it needs, and layers Guardrails, Knowledge Bases, or agents where the product calls for them.
What is Bedrock, in practical terms?
Bedrock is a managed AWS service that hosts foundation models from different providers. Instead of one contract and key per vendor, your workloads call Bedrock endpoints in an AWS Region and use the identity and audit tools already in your account.
- Chat, summarizing, and extraction inside an AWS-hosted application.
- Trying different providers' models behind one request shape.
- Document question answering through managed Knowledge Bases.
- Content and privacy rules enforced by Guardrails on model input and output.
- Agents that call Lambda functions or other AWS services.
Claude is one of the model families offered; the Claude API integration page covers direct use. For a wider comparison, see LLM integration.
Converse, InvokeModel, or something else?
InvokeModel is the low-level call: you send the model-specific JSON body and parse the model-specific reply. Converse is a higher-level API with one message structure across many models, and it handles tool use and streaming in a consistent way. If you may switch models, Converse saves rework.
Calls go through the AWS SDK to the Bedrock runtime in a Region. Permission comes from an IAM role on your compute, such as a Lambda execution role or an ECS task role, with a policy naming the invoke actions and the model resources allowed. Avoid wildcard policies. Account-level model access settings have changed over time, so check what AWS asks of you today.
Traffic can stay on AWS's network through VPC endpoints, and calls appear in CloudTrail and your existing monitoring.
Where do Guardrails and Knowledge Bases come in?
Guardrails are policies you define once, such as denied topics, word filters, or masking of personal data, and attach to model calls. They can also apply to calls made by Knowledge Bases and agents. Your app must still decide what to show when a guardrail intervenes.
Knowledge Bases manage ingestion from storage such as S3, chunking, embedding, a vector store, and retrieval, so you can ask questions about company files without building the pipeline. You can also skip them and run your own retrieval over a vector database for finer control.
| Layer | Use it when | Skip it when |
|---|---|---|
| Guardrails | Policy must be enforced centrally | A single narrow prompt suffices |
| Knowledge Bases | Files in S3 and you want managed retrieval | You need custom chunking or ranking |
| Agents | A task needs several AWS actions in sequence | One model call does the job |
What increases the effort on AWS?
The typical effort shown above covers a single model call path. Plan extra time for:
- Separate AWS accounts for test and production, and roles in each.
- Region selection, since model availability differs and data may need to stay in one place.
- Throughput limits and how the app behaves when it is throttled.
- Designing guardrail policies and the user-facing message for a blocked reply.
- Choosing managed Knowledge Bases or custom retrieval.
- Tagging calls so spend can be split by team or feature.
Rates differ by provider. Compare on LLM pricing and see AI LLM integration cost.
What are the usual Bedrock mistakes?
- AccessDenied because the policy names the wrong model resource or Region.
- Writing to the InvokeModel body format of one provider, then failing when the model changes.
- Granting full Bedrock access to a role that only needs to invoke one model.
- Throttling under launch traffic because nobody checked limits.
- Assuming each model supports the same features, like tool use or image input.
- A guardrail blocks a legitimate answer and the user sees an empty screen.
Is a no-code connector a fit for Bedrock?
Make and Zapier can reach models for simple automations. Because most Bedrock projects exist for IAM and network control, they tend to be custom builds. See iPaaS.
| Need | Connector | Custom build |
|---|---|---|
| Quick internal trial | Acceptable | More than needed |
| Role-based access per service | Shared credentials | IAM role per workload |
| Traffic kept off the public internet | Not possible | VPC endpoints |
| Swap models without rewriting | Tool dependent | Converse API layer |
What should you test before launch?
- Reproduce production IAM and networking in a test account.
- Call every chosen model from the real role and Region.
- Score outputs against a set of real examples after every prompt change.
- Force throttling, a guardrail block, a timeout, and malformed output.
- Run a load test and set alarms at your limits.
- Pilot with a small group and CloudTrail logging on.
- Track quality and spend, with the earlier release ready to restore.
How does StackLift scope a Bedrock integration?
StackLift integrates with Bedrock as an independent engineering partner and has no affiliation with AWS. A senior architect reviews account layout, IAM, and Region choices before scope is treated as a commitment, and the free estimate prices features one by one.
Agents stay on feature branches and never handle production secrets. CI/CD, observability, security review, and rollback are part of the plan from the first sprint.
Questions people ask
What does Converse do that InvokeModel does not?
It gives one request and response shape across many models, including tool use and streaming, so swapping models needs less code.
Does Bedrock need provider API keys?
Your code authenticates to AWS, normally with an IAM role. You do not hold a separate key from each model provider.
Do Guardrails work with Knowledge Bases?
AWS describes them as applying to model invocations, including those made by Knowledge Bases and agents. Confirm the current behavior in the docs.
Should I pick Bedrock or Vertex AI?
Usually the one matching where your workloads and data already live. Compare with the Vertex AI page, then test both on your examples.
How long does a Bedrock integration take?
Most first versions take 6 to 16 weeks depending on scope. A 3-day Scope Sprint gives you a detailed plan first.