text_similarity_reranker returns negative scores for some models

[kderusso]

Elasticsearch Version

8.16.1

Installed Plugins

No response

Java Version

bundled

OS Version

Reproducable on cloud

Problem Description

Certain supported rerank models, including cross-encoder__ms-marco-minilm-l-6-v2, return negative scores when used in conjunction with the text_similarity_reranker. Negative scores are not allowed in the query phase, so we need to handle this better.

One potential solution is linearly shifting or otherwise normalizing the returned score values, so they're always within certain parameters.

Steps to Reproduce

PUT _inference/rerank/ms-marco-minilm-l-6-v2
{
  "service": "elasticsearch",
  "task_type": "rerank",
  "service_settings": {
    "num_allocations": 1,
    "num_threads": 1,
    "model_id": "cross-encoder__ms-marco-minilm-l-6-v2"
  },
  "task_settings": {
    "return_documents": true
  }
}

PUT /my-index
{
  "mappings": {
    "properties": {
      "title": {
        "type": "text"
      }
    }
  }
}

POST /my-index/_doc/
{
  "text": "Dog training classes"
}

POST my-index/_search
{
  "retriever": {
    "text_similarity_reranker": {
      "retriever": {
        "standard": {
          "query": {
            "match": {
              "text": "dog"
            }
          }
        }
      },
      "field": "text",
      "inference_id": "ms-marco-minilm-l-6-v2",
      "inference_text": "dog",
      "rank_window_size": 100
    }
  }
}

The returned search result is:

{
  "took": 26,
  "timed_out": false,
  "_shards": {
    "total": 1,
    "successful": 1,
    "skipped": 0,
    "failed": 0
  },
  "hits": {
    "total": {
      "value": 1,
      "relation": "eq"
    },
    "max_score": -0.5682936,
    "hits": [
      {
        "_index": "my-index",
        "_id": "s1xJapQBZfij0Ahq54L5",
        "_score": -0.5682936,
        "_source": {
          "text": "Dog training classes"
        }
      }
    ]
  }
}

The score of the document is < 0.

Logs (if relevant)

No response

[elasticsearchmachine]

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

[elasticsearchmachine]

Pinging @elastic/es-search-relevance (Team:Search Relevance)

[elasticsearchmachine]

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

[elasticsearchmachine]

Pinging @elastic/search-relevance (Team:Search - Relevance)

[leemthompo]

This docs issue looks related: Clarify negative scores returned by Elastic Rerank