Calculate cosine similarity for nested object in elasticsearch using Python

[RiyaP (RIYA PAUL)]

I am working on an use case which has documents in following format:-

[{'Fruits': ['Mango', 'Apple']},
 {'Fruits': ['Banana', 'Guava', 'Mango']},
 {'Fruits': ['Grapes', 'Apple']}]

I am storing the data into Elasticsearch cluster in below format:-

["Fruits":[{"value": dense_vector('Mango'),"text": "Mango"},{"value": dense_vector('Apple'),"text": "Apple"}],
"Fruits":[{"value": dense_vector('Banana'),"text": "Banana"},{"value": dense_vector('Guava'),"text": "Guava"},
{"value": dense_vector('Mango'),"text": "Mango"}],...]

Am storing dense vector(Bert embeddings) so as to handle synonyms by calculating cosine similarity. My Elasticsearch query looks like this"-

{
    "nested": {
      "path": "Fruits",
        "score_mode": "max", 
      "query": {
        "function_score": {
          "script_score": {
            "script": {
              "source": "1+cosineSimilarity(params.query_vector0,'Fruits.value')+cosineSimilarity(params.query_vector1,'Fruits.value')",
              "params": {"query_vector0":query_vector,"query_vector1":query_vector1}
            }
          }
        }
      }
    }
}

,wherein query_vector and query_vector1 are the below values"-

query_vector=bert_embedding.encode([str(Grape)])[0]
query_vector1=bert_embedding.encode([str(Apples)])[0]

When I give user queries as "Grape" and "Apples", I am getting similar score for these 2 documents:-

[{'Fruits': ['Mango', 'Apple']},
 {'Fruits': ['Grapes', 'Apple']}]

Whereas, as expected, I should get a higher score for {'Fruits': ['Grapes', 'Apple']} when compared to the 1st document. Also , upon changing the value for "score_mode" to avg/sum/max, the results are changing. Which score_mode should I finalize on based on my document format?

Am not able to figure out how cosine score is similar for both the documents. Can someone please help on that?Preformatted text