Amazon Bedrock
Amazon Bedrock is an AWS service that lets developers call AI models from several companies through one set of APIs, with extra tools for search and safety.
Amazon Bedrock is a service from Amazon Web Services (AWS) for building apps on top of foundation models. A foundation model is a large AI model trained on huge amounts of data. Builders use it as a starting point for new apps instead of training their own model from scratch. Bedrock hosts many of them from companies including Amazon, Anthropic, DeepSeek, OpenAI and xAI. AWS runs the servers, so you do not set up or manage the hardware yourself.
The main idea is one front door. The Converse API gives chat-style models one shared request format. You send a list of messages and a model ID, and you get a reply back. Each message is marked with a role, so the model knows whether the user or the model said it. A model ID is a short name that points to one exact model. To try another model, you usually swap the ID and keep the rest of your code.
Two add-ons do much of the practical work. A knowledge base uses RAG (retrieval-augmented generation). That means it looks up relevant facts before the model writes. When a query arrives, Bedrock searches your data for matching information and passes it to the model. With a managed knowledge base, AWS runs the storage and search for you. In that case, you connect sources such as Amazon S3, SharePoint or Google Drive. Citations let people compare an answer with the original document and check whether it is accurate.
Guardrails are the other add-on. Content filters catch harmful text or images. A grounding check tries to spot replies that your source documents do not back up. A denied topic is a subject you do not want your app to discuss. Picture a bank that wants a chatbot for its customers. AWS uses this exact case as a guardrail example. The guardrail blocks questions or replies about illegal investment advice. If a denied topic is detected, the guardrail sends back a blocked message instead.
For the Converse API, AWS says Bedrock does not store the text, images or documents you send. They are used only to make the reply. Because the Converse API does not store the chat, your app must resend the earlier messages each turn to keep context. AWS says the model makers cannot see what customers type or what the model replies.
Building an app on top of AI models raises three awkward questions.
Follow one question from an app through Bedrock.
- 1 · sendYour app sends a message to Bedrock and names the model it wants with a model ID.
- 2 · checkIf you attached a guardrail, it screens the user's input for harmful content or private details.
- 3 · retrieveIf you use a knowledge base, Bedrock searches your own documents and adds useful passages to the prompt.
- 4 · generateThe chosen model writes a reply, whichever provider made it.
- 5 · returnThe guardrail can screen the reply too, and the answer goes back to your app.
The model ID is usually the main part you change to try a different maker's model, though some models need their own extra settings.
| Who | What they ask | What it works with |
|---|---|---|
| Support team | “Answer customer questions from our help-centre articles.” | A knowledge base built from the help-centre pages |
| Bank app team | “Stop the chatbot from giving investment advice.” | A guardrail with a denied topic for investment advice |
| Online shop | “Hide customers' personal details in chatbot questions and replies.” | A guardrail filter that masks personal data |
| Developer | “Try the same chat feature on two different models.” | The Converse API with a different model ID |
- It gives access to many foundation models from different companies through one managed service.
- The Converse API lets one piece of chat code work across many models.
- Knowledge bases let a model answer with your own documents and include citations.
- Guardrails can filter harmful content and mask personal details in inputs and replies.
- A guardrail catches much but not all; its detection of sensitive information is probabilistic.
- Some models still need their own special settings, even through the shared API.
- Answers still need checking; citations let you compare each answer with the original document.
- It runs on AWS, so you still need an AWS account and permissions to call it.
Sources used
This explainer is written in original language. The links below support its factual claims.
- docsWhat is Amazon Bedrock?, Amazon Web Services · read 29 Sept 2026
- docsInference using Converse API, Amazon Web Services · read 29 Sept 2026
- docsRetrieve data and generate AI responses with Amazon Bedrock Knowledge Bases, Amazon Web Services · read 29 Sept 2026
- docsDetect and filter harmful content by using Amazon Bedrock Guardrails, Amazon Web Services · read 29 Sept 2026
- docsData protection in Amazon Bedrock, Amazon Web Services · read 29 Sept 2026
- officialWhat are foundation models?, Amazon Web Services · read 29 Sept 2026
- docsBlock denied topics to help remove harmful content, Amazon Web Services · read 29 Sept 2026