Link Search Menu Expand Document Documentation Menu
Documentation

Integrating ML models

OpenSearch offers support for machine learning (ML) models that you can use in conjunction with k-NN search to retrieve semantically similar documents. This semantic search capability improves search relevance for your applications.

Before you get started, you’ll need to set up and secure your cluster.

Choosing a model

To integrate an ML model into your search workflow, choose one of the following options:

  1. Local model: Upload a model to the OpenSearch cluster and use it locally. This option allows you to serve the model in your OpenSearch cluster but may require significant system resources.

    1. Pretrained model provided by OpenSearch: This option requires minimal setup and avoids the time and effort required to train a custom model.

      For a list of supported models and information about using a pretrained model provided by OpenSearch, see Pretrained models.

    2. Custom model: This option offers customization for your specific use case.

      For information about uploading your model, see Using ML models within OpenSearch.

  2. Externally hosted model: This option allows you to connect to a model hosted on a third-party platform. It requires more setup but allows the use of models that are already hosted on a service other than OpenSearch.

    To connect to an externally hosted model, you need to set up a connector:

In OpenSearch version 2.9 and later, you can integrate local and external models simultaneously within a single cluster.

Tutorial

For a step-by-step tutorial, see Neural search tutorial.

Using a model

You can use an ML model in one of the following ways:

Invoking a model for inference

You can invoke your model by calling the Predict API. For example, testing text embedding models lets you see the vector embeddings they generate.

Models trained through the ML Commons plugin support model-based algorithms, such as k-means. After you’ve trained a model to your precision requirements, you can use such a model for inference. Alternatively, you can use the Train and Predict API to test your model without having to evaluate the model’s performance.

OpenSearch supports multiple search methods that integrate with ML models. For more information, see Search methods.

Disabling a model

You can temporarily disable a model when you don’t want to undeploy or delete it. Disable a model by calling the Update Model API and setting is_enabled to false. When you disable a model, it becomes unavailable for Predict API requests. If you disable a model that is undeployed, the model remains disabled after deployment. You’ll need to enable it in order to use it for inference.

Rate limiting inference calls

Setting a rate limit for Predict API calls on your ML models allows you to reduce your model inference costs. You can set a rate limit for the number of Predict API calls at the following levels:

  • Model level: Configure a rate limit for all users of the model by calling the Update Model API and specifying a rate_limiter. For more information, see Update Model API.
  • User level: Configure a rate limit for a specific user or users of the model by creating a controller. A model may be shared by multiple users; you can configure the controller to set different rate limits for different users. For more information, see Create Controller API.

Model-level rate limiting applies to all users of the model. If you specify both a model-level rate limit and a user-level rate limit, the overall rate limit is set to the more restrictive of the two. For example, if the model-level limit is 2 requests per minute and the user-level limit is 4 requests per minute, the overall limit will be set to 2 requests per minute.

To set the rate limit, you must provide two inputs: the maximum number of requests and the time frame. OpenSearch uses these inputs to calculate the rate limit as the maximum number of requests divided by the time frame. For example, if you set the limit to be 4 requests per minute, the rate limit is 4 requests / 1 minute, which is 1 request / 0.25 minutes, or 1 request / 15 seconds. OpenSearch processes predict requests sequentially, in a first-come-first-served manner, and will limit those requests to 1 request per 15 seconds. Imagine two users, Alice and Bob, calling the Predict API for the same model, which has a rate limit of 1 request per 15 seconds. If Alice calls the Predict API and immediately after that Bob calls the Predict API, OpenSearch processes Alice’s predict request and rejects Bob’s request. Once 15 seconds has passed since Alice’s request, Bob can send a request again, and this request will be processed.

Related articles

350 characters left

Have a question? .

Want to contribute? or .