Cosine-Distance of reduced dense vectors

[christiana]

Hi,

i have set up a database containing documents which are having dense vectors for the features and for a binary [1,0] mask.
I would like to use the script query to the database with a feature vector at runtime. For this i'm using cosine-similarity, but till now only between the entire feature vectors. But when querying i want to filter the features with a mask vector to get only a subset of entries. Then I want to calculate the cosine-similarity.

My db mapping looks like:

mapping = {
        "mappings": {
            "properties": {
                   "imageid": {
                            "type": "text"
                    },
                     "score": {
                            "type": "float"
                     },
                      'feature': {
                              "type": "dense_vector",
                              "dims": dim
                       }
                      'maskvec': {
                              "type": "dense_vector",
                              "dims": dim
                       }
            }               
        }
    }

The script score query till now looks like this:

"script_score": {
                            "query": {
                                "match_all": {}
                            },
                            "script": {
                                "lang":"painless",
                                "source": """
                                return cosineSimilarity(params.queryVector, feature) + 1.0
                                """,
                                "params": {
                                    "queryVector": list(featurevector)
                                }    
                            }
                        }

I would like to do sth like this:
mask = params.querymask and maskvec
queryvec = params.queryVector[mask]
featurevec = feature[mask]
return cosineSimilarity(queryvec, featurevec) + 1.0

Is it possible to implement this with elastic search?
Cheers,
Christian