[ML] Add prefix strings option to trained models

[davidkyle]

Certain NLP models such as multilingual-e5-large require a prefix string to be applied to the input text. For asymmetric tasks such as information retrieval the prefix can be different when ingesting the data and when searching it. For example text embedding model can have a one prefix applied when the model is evaluated as part of an knn search and a different prefix when ingesting documents.

An example model configuration with prefix strings is:

{
    "model_type": "pytorch",
    "inference_config": {
        "text_embedding": {
            "tokenization": {
                "xlm_roberta": {
                    "do_lower_case": false,
                    "with_special_tokens": true,
                    "max_sequence_length": 512,
                    "truncate": "first",
                    "span": -1
                }
            },
            "embedding_size": 384
        }
    },
    "prefix_strings": {
        "search": "this is a query",
        "ingest": "this is a passage"
    }
}

Many files have been touched by this change but the bulk of the work is quite simple: define the configuration object an pass the prefix type (context) parameter down the inference calls.

[elasticsearchmachine]

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

[elasticsearchmachine]

Pinging @elastic/ml-core (Team:ML)

[droberts195]

LGTM, but if we're going to have packaged models on GCS that use this feature then the new section will need to also go in this class:

elasticsearch/x-pack/plugin/core/src/main/java/org/elasticsearch/xpack/core/ml/inference/trainedmodel/ModelPackageConfig.java

Line 33 in 8d6ded3

public class ModelPackageConfig implements ToXContentObject, Writeable {

[szabosteve]

Docs LGTM, thanks!

[davidkyle]

@elasticmachine update branch

[elasticmachine]

merge conflict between base and head

[davidkyle]

Closing as I've made a mess with git. #102089 is raised as a replacement