Sparse vector query

The sparse vector query executes a query consisting of sparse vectors, such as built by a learned sparse retrieval model. This can be achieved with one of two strategies:

• Using an natural language processing model to convert query text into a list of token-weight pairs
• Sending in precalculated token-weight pairs as query vectors

These token-weight pairs are then used in a query against a sparse vector or a semantic_text field with a compatible sparse inference model. At query time, query vectors are calculated using the same inference model that was used to create the tokens. When querying, these query vectors are ORed together with their respective weights, which means scoring is effectively a dot product calculation between stored dimensions and query dimensions.

For example, a stored vector {"feature_0": 0.12, "feature_1": 1.2, "feature_2": 3.0} with query vector {"feature_0": 2.5, "feature_2": 0.2} would score the document _score = 0.12*2.5 + 3.0*0.2 = 0.9

 GET _search {
   "query":{
      "sparse_vector": {
        "field": "ml.tokens",
        "inference_id": "the inference ID to produce the token weights",
        "query": "the query string"
      }
   }
}

 GET _search {
   "query":{
      "sparse_vector": {
        "field": "ml.tokens",
        "query_vector": { "token1": 0.5, "token2": 0.3, "token3": 0.2 }
      }
   }
}

field

(Required, string) The name of the field that contains the token-weight pairs to be searched against.

inference_id

(Optional, string) The inference ID to use to convert the query text into token-weight pairs. It must be the same inference ID that was used to create the tokens from the input text. Only one of inference_id and query_vector is allowed. If inference_id is specified, query must also be specified. If all queried fields are of type semantic_text, the inference ID associated with the semantic_text field will be inferred. You can reference a deployment_id of a machine learning trained model deployment as an inference_id. For example, if you download and deploy the ELSER model in the Machine learning trained models UI in Kibana, you can use the deployment_id of that deployment as the inference_id.

query

(Optional, string) The query text you want to use for search. If inference_id is specified, query must also be specified. If query_vector is specified, query must not be specified.

query_vector

(Optional, dictionary) A dictionary of token-weight pairs representing the precomputed query vector to search. Searching using this query vector will bypass additional inference. Only one of inference_id and query_vector is allowed.

prune

(Optional, boolean) [preview] Whether to perform pruning, omitting the non-significant tokens from the query to improve query performance. If prune is true but the pruning_config is not specified, pruning will occur but default values will be used. Default: false.

pruning_config

(Optional, object) [preview] Optional pruning configuration. If enabled, this will omit non-significant tokens from the query in order to improve query performance. This is only used if prune is set to true. If prune is set to true but pruning_config is not specified, default values will be used.

Parameters for pruning_config are:

tokens_freq_ratio_threshold

(Optional, integer) [preview] Tokens whose frequency is more than tokens_freq_ratio_threshold times the average frequency of all tokens in the specified field are considered outliers and pruned. This value must between 1 and 100. Default: 5.

tokens_weight_threshold

(Optional, float) [preview] Tokens whose weight is less than tokens_weight_threshold are considered insignificant and pruned. This value must be between 0 and 1. Default: 0.4.

only_score_pruned_tokens

(Optional, boolean) [preview] If true we only input pruned tokens into scoring, and discard non-pruned tokens. It is strongly recommended to set this to false for the main query, but this can be set to true for a rescore query to get more relevant results. Default: false.

Note

The default values for tokens_freq_ratio_threshold and tokens_weight_threshold were chosen based on tests using ELSERv2 that provided the most optimal results.

The following is an example of the sparse_vector query that references the ELSER model to perform semantic search. For a more detailed description of how to perform semantic search by using ELSER and the sparse_vector query, refer to this tutorial.

 GET my-index/_search {
   "query":{
      "sparse_vector": {
         "field": "ml.tokens",
         "inference_id": "my-elser-model",
         "query": "How is the weather in Jamaica?"
      }
   }
}

Multiple sparse_vector queries can be combined with each other or other query types. This can be achieved by wrapping them in boolean query clauses and using linear boosting:

 GET my-index/_search {
  "query": {
    "bool": {
      "should": [
        {
          "sparse_vector": {
            "field": "ml.inference.title_expanded.predicted_value",
            "inference_id": "my-elser-model",
            "query": "How is the weather in Jamaica?",
            "boost": 1
          }
        },
        {
          "sparse_vector": {
            "field": "ml.inference.description_expanded.predicted_value",
            "inference_id": "my-elser-model",
            "query": "How is the weather in Jamaica?",
            "boost": 1
          }
        },
        {
          "multi_match": {
            "query": "How is the weather in Jamaica?",
            "fields": [
              "title",
              "description"
            ],
            "boost": 4
          }
        }
      ]
    }
  }
}

This can also be achieved using reciprocal rank fusion (RRF), through an rrf retriever with multiple standard retrievers.

 GET my-index/_search {
  "retriever": {
    "rrf": {
      "retrievers": [
        {
          "standard": {
            "query": {
              "multi_match": {
                "query": "How is the weather in Jamaica?",
                "fields": [
                  "title",
                  "description"
                ]
              }
            }
          }
        },
        {
          "standard": {
            "query": {
              "sparse_vector": {
                "field": "ml.inference.title_expanded.predicted_value",
                "inference_id": "my-elser-model",
                "query": "How is the weather in Jamaica?",
                "boost": 1
              }
            }
          }
        },
        {
          "standard": {
            "query": {
              "sparse_vector": {
                "field": "ml.inference.description_expanded.predicted_value",
                "inference_id": "my-elser-model",
                "query": "How is the weather in Jamaica?",
                "boost": 1
              }
            }
          }
        }
      ],
      "window_size": 10,
      "rank_constant": 20
    }
  }
}

The following is an extension to the above example that adds a [preview] pruning configuration to the sparse_vector query. The pruning configuration identifies non-significant tokens to prune from the query in order to improve query performance.

Token pruning happens at the shard level. While this should result in the same tokens being labeled as insignificant across shards, this is not guaranteed based on the composition of each shard. Therefore, if you are running sparse_vector with a pruning_config on a multi-shard index, we strongly recommend adding a Rescore filtered search results function with the tokens that were originally pruned from the query. This will help mitigate any shard-level inconsistency with pruned tokens and provide better relevance overall.

 GET my-index/_search {
   "query":{
      "sparse_vector":{
         "field": "ml.tokens",
         "inference_id": "my-elser-model",
         "query":"How is the weather in Jamaica?",
         "prune": true,
         "pruning_config": {
           "tokens_freq_ratio_threshold": 5,
           "tokens_weight_threshold": 0.4,
           "only_score_pruned_tokens": false
         }
      }
   },
   "rescore": {
      "window_size": 100,
      "query": {
         "rescore_query": {
            "sparse_vector": {
               "field": "ml.tokens",
               "inference_id": "my-elser-model",
               "query": "How is the weather in Jamaica?",
               "prune": true,
               "pruning_config": {
                   "tokens_freq_ratio_threshold": 5,
                   "tokens_weight_threshold": 0.4,
                   "only_score_pruned_tokens": true
               }
            }
         }
      }
   }
}