Fix bug where intercepted semantic knn queries did not respect filters

[kderusso]

Fixes a bug where knn queries on semantic_text did not respect configured filters.

Example query to test:

GET my-index-1/_search
{
  "query": {
    "knn": {
      "field": "inference_field",
      "k": 10,
      "num_candidates": 100,
      "query_vector_builder": {
        "text_embedding": {
          "model_id": "my-e5-model",
          "model_text": "test"
        }
      },
      "filter": {
        "term": { "text_field": "foo" } 
      }
    }
  }
}

[elasticsearchmachine]

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

[jimczi]

LGTM

[Mikep86]

Good catch and nice fix!

[Samiul-TheSoccerFan]

LGTM, left some questions to gain more knowledge.

[Samiul-TheSoccerFan]

is this value semantic_knn_filter_fix defined somewhere, or does the NodeFeature require a unique string to identify the fix from a specific version?

[kderusso]

It's required for testing, so we only test the fix on versions that have the feature. This is the pattern we use for defining node features.

[Samiul-TheSoccerFan]

Curious—how do we ensure that the same filter is not added twice? Also, do we handle a different type of filter using copy.addFilterQuery(new TermsQueryBuilder(IndexFieldMapper.NAME, indices))?

[kderusso]

This is adding the filters that were already added to the request, we do not need to do further analysis as we're translating the query. Adding the filter on index name allows us to filter to only those indices we want to query.

[Samiul-TheSoccerFan]

Adding the filter on index name allows us to filter to only those indices we want to query I missed this key point. Thank you for the clarifications.

[elasticsearchmachine]

💚 Backport successful

Status	Branch	Result
✅	8.18
✅	8.x
✅	9.0