Configure your model provider - Amazon SageMaker
Configuration in the chat UIConfiguration in a notebook
Services or capabilities described in Amazon Web Services documentation might vary by Region. To see the differences applicable to the China Regions, see Getting Started with Amazon Web Services in China (PDF).

Configure your model provider

Note

In this section, we assume that the language and embedding models that you plan to use are already deployed. For models provided by Amazon, you should already have the ARN of your SageMaker endpoint or access to Amazon Bedrock. For other model providers, you should have the API key used to authenticate and authorize requests to your model.

Jupyter AI supports a wide range of model providers and language models, refer to the list of its supported models to stay updated on the latest available models. For information on how to deploy a model provided by JumpStart, see Deploy a Model in the JumpStart documentation. You need to request access to Amazon Bedrock to use it as your model provider.

The configuration of Jupyter AI varies depending on whether you are using the chat UI or magic commands.

Configure your model provider in the chat UI

Note

You can configure several LLMs and embedding models following the same instructions. However, you must configure at least one Language model.

To configure your chat UI

  1. In JupyterLab, access the chat interface by choosing the chat icon in the left navigation panel.

  2. Choose the configuration icon in the top right corner of the left pane. This opens the Jupyter AI configuration panel.

  3. Fill out the fields related to your service provider.

    • For models provided by JumpStart or Amazon Bedrock

      • In the language model dropdown list, select sagemaker-endpoint for models deployed with JumpStart or bedrock for models managed by Amazon Bedrock.

      • The parameters differ based on whether your model is deployed on SageMaker or Amazon Bedrock.

        • For models deployed with JumpStart:

          • Enter the name of your endpoint in Endpoint name, and then the Amazon Web Services Region in which your model is deployed in Region name. To retrieve the ARN of the SageMaker endpoints, navigate to https://console.amazonaws.cn/sagemaker/ and then choose Inference and Endpoints in the left menu.

          • Paste the JSON of the Request schema tailored to your model, and the corresponding Response path for parsing the model's output.

            Note

            You can find the request and response format of various of JumpStart foundation models in the following example notebooks. Each notebook is named after the model it demonstrates.

        • For models managed by Amazon Bedrock: Add the Amazon profile storing your Amazon credentials on your system (optional), and then the Amazon Web Services Region in which your model is deployed in Region name.

      • (Optional) Select an embedding model to which you have access. Embedding models are used to capture additional information from local documents, enabling the text generation model to respond to questions within the context of those documents.

      • Choose Save Changes and navigate to the left arrow icon in the top left corner of the left pane. This opens the Jupyter AI chat UI. You can start interacting with your model.

    • For models hosted by third-party providers

      • In the language model dropdown list, select your provider ID. You can find the details of each provider, including their ID, in Jupyter AI list of model providers.

      • (Optional) Select an embedding model to which you have access. Embedding models are used to capture additional information from local documents, enabling the text generation model to respond to questions within the context of those documents.

      • Insert your models' API keys.

      • Choose Save Changes and navigate to the left arrow icon in the top left corner of the left pane. This opens the Jupyter AI chat UI. You can start interacting with your model.

The following snapshot is an illustration of the chat UI configuration panel set to invoke a Flan-t5-small model provided by JumpStart and deployed in SageMaker.

Chat UI configuration panel set to invoke a Flan-t5-small model provided by JumpStart.

Pass extra model parameters and custom parameters to your request

Your model may need extra parameters, like a customized attribute for user agreement approval or adjustments to other model parameters such as temperature or response length. We recommend configuring these settings as a start up option of your JupyterLab application using a Lifecycle Configuration. For information on how to create a Lifecycle Configuration and attach it to your domain, or to a user profile from the SageMaker console, see Create and associate a lifecycle configuration. You can choose your LCC script when creating a space for your JupyterLab application.

Use the following JSON schema to configure your extra parameters:

{
  "AiExtension": {
    "model_parameters": {
      "<provider_id>:<model_id>": { Dictionary of model parameters which is unpacked and passed as-is to the provider.}
      }
    }
  }
}

The following script is an example of a JSON configuration file that you can use when creating a JupyterLab application LCC to set the maximum length of an AI21 Labs Jurassic-2 model deployed on Amazon Bedrock. Increasing the length of the model's generated response can prevent the systematic truncation of your model's response.

#!/bin/bash
set -eux

mkdir -p /home/sagemaker-user/.jupyter

json='{"AiExtension": {"model_parameters": {"bedrock:ai21.j2-mid-v1": {"model_kwargs": {"maxTokens": 200}}}}}'
# equivalent to %%ai bedrock:ai21.j2-mid-v1 -m {"model_kwargs":{"maxTokens":200}}

# File path
file_path="/home/sagemaker-user/.jupyter/jupyter_jupyter_ai_config.json"

#jupyter --paths

# Write JSON to file
echo "$json" > "$file_path"

# Confirmation message
echo "JSON written to $file_path"

restart-jupyter-server

# Waiting for 30 seconds to make sure the Jupyter Server is up and running
sleep 30

The following script is an example of a JSON configuration file for creating a JupyterLab application LCC used to set additional model parameters for an Anthropic Claude model deployed on Amazon Bedrock.

#!/bin/bash
set -eux

mkdir -p /home/sagemaker-user/.jupyter

json='{"AiExtension": {"model_parameters": {"bedrock:anthropic.claude-v2":{"model_kwargs":{"temperature":0.1,"top_p":0.5,"top_k":25
0,"max_tokens_to_sample":2}}}}}'
# equivalent to %%ai bedrock:anthropic.claude-v2 -m {"model_kwargs":{"temperature":0.1,"top_p":0.5,"top_k":250,"max_tokens_to_sample":2000}}

# File path
file_path="/home/sagemaker-user/.jupyter/jupyter_jupyter_ai_config.json"

#jupyter --paths

# Write JSON to file
echo "$json" > "$file_path"

# Confirmation message
echo "JSON written to $file_path"

restart-jupyter-server

# Waiting for 30 seconds to make sure the Jupyter Server is up and running
sleep 30

Once you have attached your LCC to your domain, or user profile, add your LCC to your space when launching your JupyterLab application. To ensure that your configuration file is updated by the LCC, run more ~/.jupyter/jupyter_jupyter_ai_config.json in a terminal. The content of the file should correspond to the content of the JSON file passed to the LCC.

Configure your model provider in a notebook

To invoke a model via Jupyter AI within JupyterLab or Studio Classic notebooks using the %%ai and %ai magic commands

  1. Install the client libraries specific to your model provider in your notebook environment. For example, when using OpenAI models, you need to install the openai client library. You can find the list of the client libraries required per provider in the Python package(s) column of the Jupyter AI Model providers list.

    Note

    For models hosted by Amazon, boto3 is already installed in the SageMaker Distribution image used by JupyterLab, or any Data Science image used with Studio Classic.

    • For models hosted by Amazon

      Ensure that your execution role has the permission to invoke your SageMaker endpoint for models provided by JumpStart or that you have access to Amazon Bedrock.

    • For models hosted by third-party providers

      Export your provider's API key in your notebook environment using environment variables. You can use the following magic command. Replace the provider_API_key in the command by the environment variable found in the Environment variable column of the Jupyter AI Model providers list for your provider.

      %env provider_API_key=your_API_key