Elastic Inference Service

Serverless Stack Self-Managed Unavailable

Elastic Inference Service (EIS) enables you to leverage AI-powered search as a service without deploying a model in your environment. With EIS, you don't need to manage the infrastructure and resources required for machine learning inference by adding, configuring, and scaling machine learning nodes. Instead, you can use machine learning models for ingest, search, and chat independently of your Elasticsearch infrastructure.

• Your Elastic deployment or project comes with an Elastic Managed LLM connector with a default General Purpose LLM. This connector is used in Agent Builder, the AI Assistant, Attack Discovery, Automatic Import and Search Playground. For the list of available models, refer to the documentation.

• You can use ELSER to perform semantic search as a service (ELSER on EIS). Stack GA 9.2.0 Serverless

• You can use the jina-embeddings-v3 multilingual dense vector embedding model to perform semantic search through the Elastic Inference Service. Stack Planned Serverless Preview

Requests through the Elastic Managed LLM are currently proxying to AWS Bedrock in AWS US regions, beginning with us-east-1. The request routing does not restrict the location of your deployments.

ELSER requests are managed by Elastic's own EIS infrastructure and are also hosted in AWS US regions, beginning with us-east-1. All Elastic Cloud hosted deployments and serverless projects in any CSP and region can access the endpoint. As we expand the service to Azure, GCP, and more regions, we will automatically route requests to the same CSP and the closest region where the Elasticsearch cluster is hosted.

Serverless Stack GA 9.2.0

ELSER on EIS enables you to use the ELSER model on GPUs, without having to manage your own ML nodes. We expect better performance for ingest throughput than ML nodes and equivalent performance for search latency. We will continue to benchmark, remove limitations and address concerns.

You can now use semantic_text with the new ELSER endpoint on EIS. To learn how to use the .elser-2-elastic inference endpoint, refer to Using ELSER on EIS.

Semantic Search with semantic_text has a detailed tutorial on using the semantic_text field and using the ELSER endpoint on EIS instead of the default endpoint. This is a great way to get started and try the new endpoint.

Batches are limited to a maximum of 16 documents. This is particularly relevant when using the _bulk API for data ingestion.

Serverless Preview Stack Planned

You can use the jina-embeddings-v3 model through Elastic Inference Service. Running the model on EIS means that you use the model on GPUs, without the need of managing infrastructure and model resources.

Create an inference endpoint that references the jina-embeddings-v3 model in the model_id field.

				PUT _inference/text_embedding/eis-jina-embeddings-v3
					{
  "service": "elastic",
  "service_settings": {
    "model_id": "jina-embeddings-v3"
  }
}
		

The created inference endpoint uses the model for inference operations on the Elastic Inference Service. You can reference the inference_id of the endpoint in index mappings for the semantic_text field type, text_embedding inference tasks, or search queries.

The service enforces rate limits on an ongoing basis. Exceeding a limit results in HTTP 429 responses from the server until the sliding window moves on further and parts of the limit resets.

All models on EIS incur a charge per million tokens. The pricing details are available on our Pricing page.

This pricing model differs from the existing Machine Learning Nodes, which is billed through VCUs consumed.

EIS is billed per million tokens used:

• For chat models, input and output tokens are billed. Longer conversations with extensive context or detailed responses will consume more tokens.
• For embeddings models, only input tokens are billed.

Tokens are the fundamental units that language models process for both input and output. Tokenizers convert text into numerical data by segmenting it into subword units. A token can be a complete word, part of a word, or a punctuation mark, depending on the model's trained tokenizer and the frequency patterns in its training data.

For example, the sentence It was the best of times, it was the worst of times. contains 52 characters but would tokenize into approximately 14 tokens with a typical word-based approach, though the exact count varies by tokenizer.

To track your token consumption:

1. Navigate to Billing and subscriptions > Usage in the Elastic Cloud Console.
2. Look for line items where the Billing dimension is set to "Inference".