Inaccurate CreatedAt returned by ToChatResponse

[verdie-g]

ChatResponseUpdate[] updates =
[
    new()
    {
        Role = ChatRole.Assistant,
        Contents = [new TextContent("Hi")],
        CreatedAt = DateTimeOffset.UtcNow,
    },
    new()
    {
        Role = ChatRole.Assistant,
        Contents = [new UsageContent { Details = new UsageDetails { TotalTokenCount = 1 } }],
        CreatedAt = DateTimeOffset.UtcNow + TimeSpan.FromMinutes(11),
    },
];
Console.WriteLine(updates.ToChatResponse().Messages[0].CreatedAt);

In this example Messages[0].CreatedAt will get the value of the second update. I think it would make more sense to get the value of the first update as the second update is not really part of the message.

In the mistral API, I noticed significant delay between the usage update and the last text update so I would like it to be reflected in the ChatResponse.

[stephentoub]

If you do non-streaming with a really large response, is the creation date you get back closer to the time of the request or the time of completion?

[verdie-g]

python -c 'from datetime import datetime; print(datetime.now().strftime("%Y-%m-%d %H:%M:%S.%f"))'; curl -s https://api.mistral.ai/v1/conversations \
 -X POST \
 -H 'Content-Type: application/json' -H 'Authorization: Bearer XXX' \
 -d '                                               {
              "model": "mistral-small-latest",
              "inputs": [
                {
                  "type": "message.input",
                  "role": "user",
                  "content": [
                    {
                      "type": "text",
                      "text": "Write a science fiction novel set in a post-apocalyptic world where humans have colonized Mars. Include character development, world-building, political conflicts, and a twist ending. Use vivid descriptions and dialogue."         
                    }
                  ]
                }
              ]
              }
' | jq; python -c 'from datetime import datetime; print(datetime.now().strftime("%Y-%m-%d %H:%M:%S.%f"))'

2025-12-31 18:39:20.914023
{
  "object": "conversation.response",
  "conversation_id": "conv_019b757e292770e38731ec3b70c39950",
  "outputs": [
    {
      "object": "entry",
      "type": "message.output",
      "created_at": "2025-12-31T17:39:21.093770Z",
      "completed_at": "2025-12-31T17:39:36.196882Z",
      "id": "msg_019b757e29857434b92f1e43dc51e1e0",
      "model": "mistral-small-latest",
      "role": "assistant",
      "content": "**Title: Red Horizon**\n\n**Prologue: ..."
    }
  ],
  "usage": {
    "prompt_tokens": 45,
    "completion_tokens": 1371,
    "total_tokens": 1416
  }
}
2025-12-31 18:39:36.456108

The Mistral API actually returns both a created_at and a completed_at. I'm ignoring completed_at for the MEAI connector.

With streaming, you only get a created_at:

event: message.output.delta
data: {"type":"message.output.delta","created_at":"2025-12-31T17:44:50.912028Z","output_index":0,"id":"msg_019b758315a2708296dc282adf67de4b","content_index":0,"model":"mistral-small-latest","role":"assistant","content":"The"}

event: message.output.delta
data: {"type":"message.output.delta","created_at":"2025-12-31T17:44:50.917960Z","output_index":0,"id":"msg_019b758315a2708296dc282adf67de4b","content_index":0,"model":"mistral-small-latest","role":"assistant","content":" End"}

event: message.output.delta
data: {"type":"message.output.delta","created_at":"2025-12-31T17:44:50.926741Z","output_index":0,"id":"msg_019b758315a2708296dc282adf67de4b","content_index":0,"model":"mistral-small-latest","role":"assistant","content":".**"}

event: conversation.response.done
data: {"type":"conversation.response.done","created_at":"2025-12-31T17:44:50.939513Z","usage":{"prompt_tokens":45,"completion_tokens":1127,"total_tokens":1172}}

2025-12-31 18:44:51.135831

In the mistral API, I noticed significant delay between the usage update and the last text update so

Btw by significant I mean up to 500ms.