OpenSearch: use vector search in combination with should

Suppose my index has:

• A vector field called "text_encoded"
• A field called "field1", that can contain one or more of the following classes: "A", "B", "C" as list (ex. "field1" : ["A"], or "field1" : ["A", "C"])

I would like to write an OpenSearch query that can:

• select the samples with field1 classes containing either "A" OR "C"
• return similarity scores only

If I use the "should" instruction, it returns relevance scores instead.

The below query enables you to achieve this. Note that it uses Exact kNN with Scoring Script, as opposed to approximate kNN.

First, the script_score.query clause fetches a subset of documents where field1 is either "A" or "B". Then, the script_score.script clause performs kNN on the aforementioned set.

{
  "query": {
    "script_score": {
      "query": {
        "bool": {
          "should": [
            { "match": { "field1": "A" }},
            { "match": { "field1": "B" }}
          ]
        }
      },
      "script": {
        "source": "knn_score",
        "lang": "knn",
        "params": {
          "field": "text_encoded",
          "query_value": [0.1, 0.2, 0.3],
          "space_type": "cosinesimil"
        }
      }
    }
  }
}

Good luck!