EIS integration

[timgrein]

This PR integrates EIS (Elastic Inference Service) with Elasticsearch behind a feature flag.

Useful ES commands:

• Running Elasticsearch (inside of the ES root directory):
  • Run ES: ./gradlew run
  • Run ES in debug mode: ./gradlew run --debug-jvm (you can attach a debugger on the debug port, which will be logged)
  • Run ES and set EIS gateway URL via CLI: ./gradlew run -Dtests.es.xpack.inference.eis.gateway.url=http://localhost:8000
• Running tests:
  • Run tests inside inference API plugin: ./gradlew ':x-pack:plugin:inference:test'
  • Run a specific test class inside inference API plugin: ./gradlew ':x-pack:plugin:inference:test' --tests "org.elasticsearch.xpack.inference.external.response.elastic.ElasticInferenceServiceSparseEmbeddingsResponseEntityTests" (specify --tests and package + class name)
  • Run a specific test method inside Inference API plugin: ./gradlew ':x-pack:plugin:inference:test' --tests "org.elasticsearch.xpack.inference.external.response.elastic.ElasticInferenceServiceSparseEmbeddingsResponseEntityTests.testSparseEmbeddingsResponse_SingleEmbeddingInData_NoMeta_NoTruncation" (specify --tests and package + class + method name)
• Check style/formatting:
  • Check style in inference API plugin: ./gradlew ':x-pack:plugin:inference:checkstyleTest'
  • Apply spotless/formatting: ./gradlew spotlessApply

Testing locally:

• Start eis-model-server (or eis-gateway, if ELSERv2 endpoint is integrated) on port {PORT}
• Start ES with a configured eis-gateway: ./gradlew run -Dtests.es.xpack.inference.eis.gateway.url=http://localhost:{PORT}
• Create an EIS inference endpoint:

PUT {ES_HOST}/_inference/sparse_embedding/eis

{
    "service": "elastic",
    "service_settings": {
        "model_id": ".elser_model_2"
    }
}

• Perform inference (single embedding in a list):

POST {ES_HOST}/_inference/sparse_embedding/eis

{
    "input": "A blue sky"
}

• Perform inference (multiple embeddings):

POST {ES_HOST}/_inference/sparse_embedding/eis

{
    "input": [
        "Embed this text",
        "Embed this text, too",
        "This text should also be embedded"
    ]
}

Testing in serverless:

• Enable the feature flag for your project or a whole environment via a PR in our corresponding gitops repo via jvmOptions: "-Des.eis_feature_flag_enabled=true"
• Repeat steps from Testing locally

TODOs:

• Write tests for ElasticInferenceService
• Write tests for ElasticInferenceServiceActionCreator
• Write tests for ElasticInferenceServiceResponseHandler
• Write tests for ElasticInferenceServiceSparseEmbeddingsRequest
• Implement checkModelConfig in ElasticInferenceService
• Implement doChunkedInfer in ElasticInferenceService
• Implement truncation in ElasticInferenceServiceSparseEmbeddingsRequest
• Add docs for ElasticInferenceServiceFeature
• Handle error codes specified in inference Task API spec inside ElasticInferenceServiceResponseHandler
• (There might be some additional smaller TODOs, which I forgot here, I usually mark them with //TODO:)

When ready for review:

• Mark ready for review
• Add labels (this will ping the ML team for reviews):
  • >non-issue
  • :ml
  • Team:ML

Out of scope for this PR:

• "Always-on" experience (performing inference without creating an endpoint for service elastic upfront)
• Rate limiting
• Auth/Secret Settings

[elasticsearchmachine]

Pinging @elastic/search-inference-team (Team:Search - Inference)

[elasticsearchmachine]

Pinging @elastic/ent-search-eng (Team:SearchOrg)

[elasticsearchmachine]

Hi @timgrein, I've created a changelog YAML for you.

[davidkyle]

LGTM