Elasticsearch: Hybrid Search Optimization Tips

ELSER Limitations:

• it only supports up to 512 tokens when running text inference.
• So, if your data contains longer text excerpts that you need to be fully searchable, you are left with two options: a) use another model that supports longer text, or b) split your text into smaller segments.

 

Optimization Tip:

• 다음과 같이 Index Creation 에서 vector 필드를 _source 문서에서 제외 시키는 것.
• leverage RRF by providing the `rank` section at the end of the query

PUT my-index
{
  "mappings": {
    "_source": {
      "excludes": [
        "text_embedding.predicted_value",
        "ml.tokens"
      ]
    },
    "properties": {
      "text": {
        "type": "text"
      },
      "ml": {
        "properties": {
          "tokens": {
            "type": "rank_features"
          }
        }
      },
      "text_embedding": {
        "properties": {
          "predicted_value": {
            "type": "dense_vector",
            "dims": 384,
            "index": true,
            "similarity": "cosine"
          }
        }
      }
    }
  }
}

POST my-index/_search
{
  "_source": false,
  "fields": [ "text-field" ],
  "sub_searches": [
    {
      "query": {
        "match": {
          "text-field": "brown fox"
        }
      }
    },
    {
      "query": {
        "text_expansion": {
          "ml.tokens": {
            "model_id": ".elser_model_1",
            "model_text": "a quick brown fox jumps over a lazy dog"
          }
        }
      }
    }
  ],
  "knn": {
    "field": "image-vector",
    "query_vector": [0.1, 3.2, ..., 2.1],
    "k": 5,
    "num_candidates": 100
  },
  "rank": {
    "rrf": {}
  }
}