Query a knowledge base and retrieve data

Important

Guardrails are applied only to the input and the generated response from the LLM. They are not applied to the references retrieved from Knowledge Bases at runtime.

After your knowledge base is set up, you can query it and retrieve chunks from your source data that is relevant to the query by using the Retrieve API operation. You can also use a reranking model instead of the default Amazon Bedrock Knowledge Bases ranker to rank source chunks for relevance during retrieval.

To learn how to query your knowledge base, choose the tab for your preferred method, and then follow the steps:

Console

To test your knowledge base

Sign in to the AWS Management Console with an IAM identity that has permissions to use the Amazon Bedrock console. Then, open the Amazon Bedrock console at https://console.aws.amazon.com/bedrock.
In the left navigation pane, choose Knowledge bases.
In the Knowledge bases section, do one of the following actions:
- Choose the radio button next to the knowledge base you want to test and select Test knowledge base. A test window expands from the right.
- Choose the knowledge base that you want to test. A test window expands from the right.
In the test window, clear Generate responses for your query to return information retrieved directly from your knowledge base.
(Optional) Select the configurations icon ( ) to open up Configurations. For information about configurations, see Configure and customize queries and response generation.
Enter a query in the text box in the chat window and select Run to return responses from the knowledge base.
The source chunks are returned directly in order of relevance. Images extracted from your data source can also be returned as a source chunk.
To see details about the returned chunks, select Show source details.
- To see the configurations that you set for query, expand Query configurations.
- To view details about a source chunk, expand it by choosing the right arrow ( ) next to it. You can see the following information:
  - The raw text from the source chunk. To copy this text, choose the copy icon ( ). If you used Amazon S3 to store your data, choose the external link icon ( ) to navigate to the S3 object containing the file.
  - The metadata associated with the source chunk, if you used Amazon S3 to store your data. The attribute/field keys and values are defined in the .metadata.json file that's associated with the source document. For more information, see the Metadata and filtering section in Configure and customize queries and response generation.

Chat options

Switch to generating responses based on the retrieved source chunks by turning on Generate responses. If you change the setting, the text in the chat window will be completely cleared.
To clear the chat window, select the broom icon ( ).
To copy all the output in the chat window, select the copy icon ( ).

API

To query a knowledge base and only return relevant text from data sources, send a Retrieve request with an Agents for Amazon Bedrock runtime endpoint.

The following fields are required:

Field	Basic description
knowledgeBaseId	To specify the knowledge base to query.
retrievalQuery	Contains a `text` field to specify the query.
guardrailsConfiguration	Include guardrailsConfiguration fields such as `guardrailsId` and `guardrailsVersion` to use your guardrail with the request

The following fields are optional:

Field	Use case
nextToken	To return the next batch of responses (see response fields below).
retrievalConfiguration	To include query configurations for customizing the vector search. See KnowledgeBaseVectorSearchConfiguration for more information.

You can use a reranking model over the default Amazon Bedrock Knowledge Bases ranking model by including the rerankingConfiguration field in the KnowledgeBaseVectorSearchConfiguration. The rerankingConfiguration field maps to a VectorSearchRerankingConfiguration object, in which you can specify the reranking model to use, any additional request fields to include, metadata attributes to filter out documents during reranking, and the number of results to return after reranking. For more information, see VectorSearchRerankingConfiguration.

Note

If you the numberOfRerankedResults value that you specify is greater than the numberOfResults value in the KnowledgeBaseVectorSearchConfiguration, the maximum number of results that will be returned is the value for numberOfResults. An exception is if you use query decomposition (for more information, see the Query modifications section in Configure and customize queries and response generation. If you use query decomposition, the numberOfRerankedResults can be up to five times the numberOfResults.

The response returns the source chunks from the data source as an array of KnowledgeBaseRetrievalResult objects in the retrievalResults field. Each KnowledgeBaseRetrievalResult contains the following fields:

Field	Description
content	Contains a text source chunk in the `text` or an image source chunk in the `byteContent` field. If the content is an image, the data URI of the base64-encoded content is returned in the following format: `data:image/jpeg;base64,${base64-encoded string}`.
metadata	Contains each metadata attribute as a key and the metadata value as a JSON value that the key maps to.
location	Contains the URI or URL of the document that the source chunk belongs to.
score	The relevancy score of the document. You can use this score to analyze the ranking of results.

If the number of source chunks exceeds what can fit in the response, a value is returned in the nextToken field. Use that value in another request to return the next batch of results.

If the retrieved data contains images, the response also returns the following response headers, which contain metadata for source chunks returned in the response:

x-amz-bedrock-kb-byte-content-source – Contains the Amazon S3 URI of the image.
x-amz-bedrock-kb-description – Contains the base64-encoded string for the image.

Note

You can't filter on these metadata response headers when configuring metadata filters.

Multimodal queries

For knowledge bases using multimodal embedding models, you can query with images in addition to text. The retrievalQuery field supports a multimodalInputList field for image queries:

Note

For comprehensive guidance on setting up and working with multimodal knowledge bases, including choosing between Nova and BDA approaches, see Build a knowledge base for multimodal content.


{
    "knowledgeBaseId": "EXAMPLE123",
    "retrievalQuery": {
        "text": "Find similar shoes",
        "multimodalInputList": [
            {
                "content": {
                    "byteContent": "base64-encoded-image-data"
                },
                "modality": "IMAGE"
            }
        ]
    }
}

You can also query with images only by omitting the text field:


{
    "knowledgeBaseId": "EXAMPLE123", 
    "retrievalQuery": {
        "multimodalInputList": [
            {
                "content": {
                    "byteContent": "base64-encoded-image-data"
                },
                "modality": "IMAGE"
            }
        ]
    }
}

Common multimodal query patterns

Image-to-image search: Upload an image to find visually similar images. Example: Upload a photo of a red Nike shoe to find similar shoes in your product catalog.
Text + image refinement: Combine text and images for more precise results. Example: "Find similar shoes but with different colors" along with an uploaded shoe image.
Visual document search: Search for charts, diagrams, or visual elements within documents. Example: Upload a chart image to find similar charts in your document collection.

Choosing between Nova and BDA for multimodal content

When working with multimodal content, choose your approach based on your content type and query patterns:

Nova vs BDA Decision Matrix
Content Type	Use Nova Multimodal Embeddings	Use Bedrock Data Automation (BDA) Parser
Video Content	Visual storytelling focus (sports, ads, demonstrations), queries on visual elements, minimal speech content	Important speech/narration (presentations, meetings, tutorials), queries on spoken content, need transcripts
Audio Content	Music or sound effects identification, non-speech audio analysis	Podcasts, interviews, meetings, any content with speech requiring transcription
Image Content	Visual similarity searches, image-to-image retrieval, visual content analysis	Text extraction from images, document processing, OCR requirements

Note

Nova multimodal embeddings cannot process speech content directly. If your audio or video files contain important spoken information, use the BDA parser to convert speech to text first, or choose a text embedding model instead.

Multimodal query limitations

Maximum of one image per query in the current release
Image queries are only supported with multimodal embedding models (Titan G1 or Cohere Embed v3)
RetrieveAndGenerate API is not supported for knowledge bases with multimodal embedding models and S3 content buckets
If you provide an image query to a knowledge base using text-only embedding models, a 4xx error will be returned

Multimodal API response structure

Retrieval responses for multimodal content include additional metadata:

Source URI: Points to your original S3 bucket location
Supplemental URI: Points to the copy in your multimodal storage bucket
Timestamp metadata: Included for video and audio chunks to enable precise playback positioning

Note

When using the API or SDK, you'll need to handle file retrieval and timestamp navigation in your application. The console handles this automatically with enhanced video playback and automatic timestamp navigation.

Note

If you receive an error that the prompt exceeds the character limit while generating responses, you can shorten the prompt in the following ways:

Reduce the maximum number of retrieved results (this shortens what is filled in for the $search_results$ placeholder in the Knowledge base prompt templates: orchestration & generation).
Recreate the data source with a chunking strategy that uses smaller chunks (this shortens what is filled in for the $search_results$ placeholder in the Knowledge base prompt templates: orchestration & generation).
Shorten the prompt template.
Shorten the user query (this shortens what is filled in for the $query$ placeholder in the Knowledge base prompt templates: orchestration & generation).

Warning Javascript is disabled or is unavailable in your browser.

To use the Amazon Web Services Documentation, Javascript must be enabled. Please refer to your browser's Help pages for instructions.

Document Conventions

Test your knowledge base with queries and responses

Query a knowledge base and generate responses