Trouble implementing Elasticsearch BBQ

Elastic Stack Elasticsearch
1

I am trying to integrate BBQ, specifically bbq_hnsw into my existing index. However I'm not seeing the performance benefits I would expect.

{
        "type": "dense_vector",
        "dims": dims,
        "index": True,
        "similarity": "cosine",
        "index_options": {
            "type": "bbq_hnsw"
        }
 }

While I see that the attribute exists in the lucene _segments I see no improvement in my query time.

{
  "_shards": {
    "total": 1,
    "successful": 1,
    "failed": 0
  },
  "indices": {
    "embedding_control_20250417_qaqkj3cw": {
      "shards": {
        "0": [
          {
            "routing": {
              "state": "STARTED",
              "primary": true,
              "node": "IhQGlYPITDmtzYIEHHIwLg"
            },
            "num_committed_segments": 1,
            "num_search_segments": 1,
            "segments": {
              "_16": {
                "generation": 42,
                "num_docs": 2172,
                "deleted_docs": 0,
                "size": "35.5mb",
                "size_in_bytes": 37258972,
                "committed": true,
                "search": true,
                "version": "9.12.0",
                "compound": true,
                "attributes": {
                  "Lucene99HnswVectorsFormat": "[ibmGraniteEmbedding]",
                  "Zstd814StoredFieldsFormat.mode": "BEST_COMPRESSION"
                }
              }
            }
          }
        ]
      }
    }
  }
}

This is how my current query is setup which is currently working but I see no difference in speed when comparing a build that was embedded without the index_options. The build is 2,000 records and the model dims are 768.

"query": {
    "function_score": {
        "query": {
            "script_score": {
                "query": bool_query,
                "script": {
                    "source": f"if (doc['{self.model_info['field_name']}'].isEmpty()) {{ return 0.0; }} else {{ return cosineSimilarity(params.query_vector, '{self.model_info['field_name']}') + 1.0; }}",
                    "params": {
                        "query_vector": embedded_query
                    }
                }
            }
        },
2

You should try using a knn query instead. As you are using bbq_hnsw, that'd make more sense to me.