Cosine Similarities are not proper

[osmantamer]

Hello,

I was trying to measure the cosine similarity between vectors with dimensions (1, 128). Here is the query.

` { "query": { "function_score": { "boost_mode": "replace", "min_score": 0, "script_score": { "script": { "source": "binary_vector_score", "lang": "painless", "params": { "cosine": true, "field": "embedding_vector", "vector": [

] } } } } }, "size": 170 }` But results are meaningless. Eventhough two vectors is not similar, it's score is over 0.9. Similarity scores decrease gradually.

[durakkerem]

I am also having the same problem. Scores does not reflect actual cosine similarities.

[durakkerem]

did you resolve the problem?

[osmantamer]

did you resolve the problem?

unfortunately.

[lior-k]

Hi guys, I'm sorry for the late response. Cosine scoring works fine for us. I added a test for the scores. feel free to change it, test it & submit it to me if it fails.

@osmantamer Note that the query you pasted is wrong. it should state "lang": "knn" from the readme:

{
  "query": {
    "function_score": {
      "boost_mode": "replace",
      "script_score": {
        "script": {
          "inline": "binary_vector_score",
          "lang": "knn",
          "params": {
            "cosine": false,
            "field": "embedding_vector",
            "vector": [
               -0.09217305481433868, 0.010635560378432274, -0.02878434956073761, 0.06988169997930527, 0.1273992955684662, -0.023723633959889412, 0.05490724742412567, -0.12124507874250412, -0.023694118484854698, 0.014595639891922474, 0.1471538096666336, 0.044936809688806534, -0.02795785665512085, -0.05665992572903633, -0.2441125512123108, 0.2755320072174072, 0.11451690644025803, 0.20242854952812195, -0.1387604922056198, 0.05219579488039017, 0.1145530641078949, 0.09967200458049774, 0.2161576747894287, 0.06157230958342552, 0.10350126028060913, 0.20387393236160278, 0.1367097795009613, 0.02070528082549572, 0.19238869845867157, 0.059613026678562164, 0.014012521132826805, 0.16701748967170715, 0.04985826835036278, -0.10990987718105316, -0.12032567709684372, -0.1450948715209961, 0.13585780560970306, 0.037511035799980164, 0.04251480475068092, 0.10693439096212387, -0.08861573040485382, -0.07457160204648972, 0.0549330934882164, 0.19136285781860352, 0.03346432000398636, -0.03652812913060188, -0.1902569830417633, 0.03250952064990997, -0.3061246871948242, 0.05219300463795662, -0.07879918068647385, 0.1403723508119583, -0.08893408626317978, -0.24330253899097443, -0.07105310261249542, -0.18161986768245697, 0.15501035749912262, -0.216160386800766, -0.06377710402011871, -0.07671763002872467, 0.05360138416290283, -0.052845533937215805, -0.02905619889497757, 0.08279753476381302
             ]
          }
        }
      }
    }
  },
  "size": 100
}

[pfeiffer]

I'm also having trouble with the scoring and scoring does not seem to reflect the distance.

Did any of you manage to solve it?

I'd love to contribute with a failing test case - how can I run the tests? :-)

[osmantamer]

I'm also having trouble with the scoring and scoring does not seem to reflect the distance.

Did any of you manage to solve it?

I realize that it was my mistake. In python, compare_faces function of the face_recognition library return the euclidean distance between two faces, (smaller value means similar faces), but in this plug-in it returns cosine similarity (higher value means similar faces). That was my mistake.

[lior-k]

Glad to hear. Closing this