The threshold score does not return a result

API Bugs
1

I’m using the File Search Assistant API. I’m experiencing high request token usage, and I don’t know why. I’ve looked at the debugging steps a run follows when executing. I’ve seen that the default score_threshold is 0, so it returns all the chunks in the file with a higher threshold (all of them). Below I’ve shown the result of a run, and you can see that even if the score is very small, they are used. In the example, I limited de results to 4.

{
  "id": "call_xxxxxxxxxxxxx",
  "type": "file_search",
  "file_search": {
    "ranking_options": {
      "ranker": "default_2024_08_21",
      "score_threshold": 0
    },
    "results": [
      {
        "file_id": "file-xxxxxxxxxxxx",
        "file_name": "Onboarding.md",
        "score": 0.4028151374551781,
        "attributes": {}
      },
      {
        "file_id": "file-xxxxxxxxxxxx",
        "file_name": "Onboarding.md",
        "score": 0.3355354378068565,
        "attributes": {}
      },
      {
        "file_id": "file-xxxxxxxxxxxx",
        "file_name": "Onboarding.md",
        "score": 0.061083678706996304,
        "attributes": {}
      },
      {
        "file_id": "file-xxxxxxxxxxxx",
        "file_name": "Onboarding.md",
        "score": 0.055866281340211356,
        "attributes": {}
      }
    ]
  }
}

The solution I found was to raise the score_threshold to 0.1 and return only 2 of those 4 results, so the cost would be lower. The result was that it doesn’t return any chunks. I tried lowering the score_threshold to 0.01 and it still doesn’t return any value. I used the same file as the previous example and the same query.

{
  "id": "call_xxxxxxxxxxx",
  "type": "file_search",
  "file_search": {
    "ranking_options": {
      "ranker": "default_2024_08_21",
      "score_threshold": 0.01
    },
    "results": []
  }
}

Can anyone else confirm that they have the same problem?