Skip to main content

Use Amazon Bedrock and SageMaker Models

You can use models hosted in your own Bedrock or SageMaker account or in deepset's account.

You can use embedding models and LLMs hosted on Amazon Bedrock through Bedrock's API. For a full list of supported models, see Amazon Bedrock documentation.

Prerequisites​

To use models through your own Amazon Bedrock or SageMaker account, you must have valid AWS credentials:

  • Access key ID
  • Secret access key

For details, see Amazon Bedrock model access and Amazon Bedrock documentation.

Using Bedrock Models​

Using Bedrock Models Through deepset's Account​

If you don't have a Bedrock account, you can use LLMs hosted there through DeepsetAmazonBedrockGenerator. It uses deepset's Bedrock account, so you don't need to create your own. If you use DeepsetAmazonBedrockGenerator, you don't need to connect deepset AI Platform to Bedrock.

Using Bedrock Models Through Your Own Account​

First, connect deepset AI Platform to Amazon Bedrock by passing your Bedrock API key on the Integrations page. You can do so for a single workspace or for the whole organization:

Add Workspace-Level Integration​

  1. Click your profile icon and choose Settings.
  2. Go to Workspace>Integrations.
  3. Find the provider you want to connect and click Connect next to them.
  4. Enter the API key and any other required details.
  5. Click Connect. You can use this integration in pipelines and indexes in the current workspace.

Add Organization-Level Integration​

  1. Click your profile icon and choose Settings.
  2. Go to Organization>Integrations.
  3. Find the provider you want to connect and click Connect next to them.
  4. Enter the API key and any other required details.
  5. Click Connect. You can use this integration in pipelines and indexes in all workspaces in the current organization.

Then, add a component that uses the Bedrock model to your pipeline. Here is a list of components by the model type they use:

  • Embedding models (used to calculate embeddings for text):
    • AmazonBedrockTextEmbedder: Calculates embeddings for text, such as query. Often used in query pipelines to embed query and then pass it to an embedding retriever.
    • AmazonBedrockDocumentEmbedder: Calculates embeddings for documents. Often used in indexes to embed documents and pass them to DocumentWriter.

      Embedding Models in Query Pipelines and Indexes

      The embedding model you use to embed documents in your indexing pipeline must be the same as the embedding model you use to embed the query in your query pipeline.

      This means the embedders for your indexing and query pipelines must match. For example, if you use CohereDocumentEmbedder to embed your documents, you should use CohereTextEmbedder with the same model to embed your queries.

  • LLMs:
    • AmazonBedrockGenerator: Generates text, often used in RAG pipelines.

Usage Examples​

This is an example of how to use embedding models and an LLM hosted on Bedrock in an index and a query pipeline (each in a separate tab):

components:
  ...
     splitter:
      type: haystack.components.preprocessors.document_splitter.DocumentSplitter
      init_parameters:
        split_by: word
        split_length: 250
        split_overlap: 30

    document_embedder:
      type: haystack_integrations.components.embedders.amazon_bedrock.document_embedder.AmazonBedrockDocumentEmbedder
      init_parameters:
        model: "cohere.embed-english-v3"

    writer:
      type: haystack.components.writers.document_writer.DocumentWriter
      init_parameters:
        document_store:
          type: haystack_integrations.document_stores.opensearch.document_store.OpenSearchDocumentStore
          init_parameters:
            embedding_dim: 768
            similarity: cosine
        policy: OVERWRITE
        
connections:  # Defines how the components are connected
  ...
  - sender: splitter.documents
    receiver: document_embedder.documents
  - sender: document_embedder.documents
    receiver: writer.documents

Using SageMaker Models​

You can use LLMs hosted on SageMaker through the SagemakerGenerator component. Pass the model endpoint in the model parameter. If your model requires custom attributes, you can specify them in the aws_custom_attributes parameter. The llama2 family of models requires a custom parameter: accept_eula: True to work. See the usage example below for more details.

Usage Example​

To use an LLM hosted on SageMaker in your query pipeline, pass the model in the Generator's model parameter in Pipeline Builder:

The Generator's model parameter on the component card in Pipeline Builder

Or in the code editor:

components:
...
  prompt_builder:
      type: haystack.components.builders.prompt_builder.PromptBuilder
      init_parameters:
        template: |-
          You are a technical expert.
          You answer questions truthfully based on provided documents.
          For each document check whether it is related to the question.
          Only use documents that are related to the question to answer it.
          Ignore documents that are not related to the question.
          If the answer exists in several documents, summarize them.
          Only answer based on the documents provided. Don't make things up.
          If the documents can't answer the question or you are unsure say: 'The answer can't be found in the text'.
          These are the documents:
          {% for document in documents %}
          Document[{{ loop.index }}]:
          {{ document.content }}
          {% endfor %}
          Question: {{question}}
          Answer:

    generator:
      type: haystack_integrations.components.generators.amazon_sagemaker.sagemaker.SagemakerGenerator
      init_parameters:
         model: jumpstart-dft-hf-llm-llama2-7b-instruct-bf16 #this is the model
         generation_kwargs:
           temperature: 0.0
         aws_custom_attributes: 
           accept_eula: True
       
      ...

connections:
...
 - sender: prompt_builder.prompt
   receiver: generator.prompt
   ...
   
iputs: 
... 
   query:
   - "prompt_builder.question"
   ...
   
outputs:
...
  answers: "generator.replies"