GPT Assistant does not include resource references (tool_calls is empty)

API
1

Hi everyone!

I’m running an assistant retrieval that references specific files as a knowledgebase.
Everything works fine for subsequent messages, but the references (citations/steps/sources) don’t appear in the first message.

Does that mean the assistant replies from LLM, not from the knowledgebase?

Has anyone encountered this issue or know how to ensure these references are included from the start?

Streaming has been enabled

Check the step_details

 C:\cmd
λ curl https://api.openai.com/v1/threads/thread_twK9GX2gxbxTX2rHaMgRxr99/runs/run_UF3GBRiELVqk3elvVV59IavB/steps ^
More?   -H "Authorization: Bearer sk-proj-*******-********-lsA" ^
More?   -H "Content-Type: application/json" ^
More?   -H "OpenAI-Beta: assistants=v2"
{
  "object": "list",
  "data": [
    {
      "id": "step_9Q4FqV6VBQhuzf9oCbqkslRU",
      "object": "thread.run.step",
      "created_at": 1735224644,
      "run_id": "run_UF3GBRiELVqk3elvVV59IavB",
      "assistant_id": "asst_45S18TB8cX7rXN0GTZS2ISlm",
      "thread_id": "thread_twK9GX2gxbxTX2rHaMgRxr99",
      "type": "message_creation",
      "status": "completed",
      "cancelled_at": null,
      "completed_at": 1735224648,
      "expires_at": null,
      "failed_at": null,
      "last_error": null,
      "step_details": {
        "type": "message_creation",
        "message_creation": {
          "message_id": "msg_Z890JJ6gRtyZjpNm417oomvY"
        }
      },
      "usage": {
        "prompt_tokens": 847,
        "completion_tokens": 249,
        "total_tokens": 1096,
        "prompt_token_details": {
          "cached_tokens": 0
        }
      }
    }
  ],
  "first_id": "step_9Q4FqV6VBQhuzf9oCbqkslRU",
  "last_id": "step_9Q4FqV6VBQhuzf9oCbqkslRU",
  "has_more": false
}

While am expecting step_details something like:

"last_error": null,
      "step_details": {
        "type": "tool_calls",
        "tool_calls": [
          {
            "id": "call_1cCK2BhTX9MupMsg8gXv2hnE",
            "type": "file_search",
            "file_search": {
              "ranking_options": {
                "ranker": "default_2024_08_21",
                "score_threshold": 0.0
              },
              "results": [
                {
                  "file_id": "file-2ug1eYw7i9bFPkVZXQ216v",
                  "file_name": "نظام_العمل.txt",
                  "score": 0.8058424419236964
                },
               ....
2

It seems based on

and

There is no way to force GPT Assistant to read and reply only from the knowledge.

Please advise