Reranking results with multiple vectors

Plugins k-NN
1

I’ve created an index with two vectors with the following schema

ID, feature_vector1, feature_vector2

Here feature_vector1 is list of 500 floats and feature_vector2 is list of 100 floats
I want to have the following.

  • perform search on the entire index using feature_vector1
  • using the results of the search, I want to perform reranking of the results based on the similarity with feature_vector2 for the corresponding ID

I was able to perform KNN on feature_vector1 .
But not sure how to apply post_filter on top the result obtained on feature_vector1 . Is there a way to achieve this.

1 Like
2

Hi @navmarri,

Could you help me understand what you mean by reranking of results with feature_vector2?

Post filter would basically trim away(filter out) the documents obtained from original query. it does not rescore?

3

@vamshin

First we obtain the knn results from feature_vector1. Just for the IDs that are obtained. I want to apply knn using feaure_vector2. We can think this as chaining query performing knn on top of the results of first knn results.
Does it make sense?

4

@navmarri I see what you mean. This is more like prefilter support for k-NN which is currently not available. We are working on this feature. Support custom scoring function for vectors. Using k-NN scores in a script_score query · Issue #50 · opendistro-for-elasticsearch/k-NN · GitHub.

As a work around probably, you could do boolean and operation(intersection between results from feature_vector1 and feature_vector2). Not a complete solution but should work. You might need to provide large k for getting results from intersection.

5

Example query to work around. You can also choose the weightage for the query to reflect scoring among the matched documents

curl -X POST "localhost:9200/myindex/_search" -H 'Content-Type: application/json' -d'
{
  "size" : 2,
  "query": {
    "bool": {
      "must": [
        {
          "function_score": {
            "query": {
              "knn": {
                  "my_vector": {
                  "vector": [7, 8],
                  "k": 2
                  }       
              }
          },
            "weight": 0.5
          }
        },
        {
          "function_score": {
            "query": {
              "knn": {
                  "my_vector": {
                  "vector": [3, 4],
                  "k": 2
                  }       
              }
            },
            "weight": 0.5
          }
        }
      ]
    }
  }
}
'
6

@vamshin Thanks for the suggestion.
What is the default score_mode
Is it summation of the weight from the two functions and pick the max?

7

Yes. Its sum of the scores and picks max. This gives you ability to give more weightage to the docs from 1st knn query or 2nd knn query. In the example i mentioned, we are giving equal weightage. Note, choose a very large k(you might want to experiment), to get better results.