keyNotFound errors since v0.3.7

[CamilleScholtz]

Since the new Update I keep client.chats(query: ChatQuery()) keeps throwing errors. Example code:

       let result = try await client.chats(query: ChatQuery(
            messages: [
                .init(role: .system, content: """
            You are a music expert who knows every style, genre, artist, and album; from mainstream hits to obscure world music. You can sense any gathering's mood and craft the perfect playlist. Your job is to create a playlist that perfectly fits a shoort description we'll provide. The user will send you albums in the format `artist - title`. Return, in JSON format, those albums, in that exact same `artist - title` format, which match the given description.
            
            EXAMPLE JSON OUTPUT:
            {
                "playlist": [
                    "Philip Glass - Koyaanisqatsi",
                    "Philip Glass - Glassworks",
                    "Philip Glass - Einstein on the Beach"
                ]
            }
            
            THE DESCRIPTION: \(prompt)
            """)!,
                .init(role: .user, content: albums.map(\.description).joined(
                    separator: "\n"))!,
            ],
            model: model.rawValue,
            responseFormat: .jsonObject
        ))

This will throw:

keyNotFound(CodingKeys(stringValue: "tool_calls", intValue: nil), Swift.DecodingError.Context(codingPath: [CodingKeys(stringValue: "choices", intValue: nil), _CodingKey(stringValue: "Index 0", intValue: 0), CodingKeys(stringValue: "message", intValue: nil)], debugDescription: "No value associated with key CodingKeys(stringValue: \"tool_calls\", intValue: nil) (\"tool_calls\").", underlyingError: nil))

Using model gpt-4o. When using the DeepSeek api host/model it throws a slightly different error:

keyNotFound(CodingKeys(stringValue: "annotations", intValue: nil), Swift.DecodingError.Context(codingPath: [CodingKeys(stringValue: "choices", intValue: nil), _CodingKey(stringValue: "Index 0", intValue: 0), CodingKeys(stringValue: "message", intValue: nil)], debugDescription: "No value associated with key CodingKeys(stringValue: \"annotations\", intValue: nil) (\"annotations\").", underlyingError: nil))

Both models used to work in 0.3.6.

[nezhyborets]

Hi @CamilleScholtz . The situation with annotations is clear, I should probably make it optional for now. But the one with tool_calls... I seem to be unable to reproduce it. It would be great to see the raw json that fails the parsing. But for now to make it work I will also make tool_calls optional (even though it's declared non optional in OpenAI API reference and should be present in all responses)

[sakeven]

@nezhyborets

Use curl to test a simple call, and the result returned is that there are no tool_calls.

curl -X POST "https://api.openai.com/v1/chat/completions" \
    -H "Authorization: Bearer $OPENAI_API_KEY" \
    -H "Content-type: application/json" \
    -d '{
        "model": "gpt-4o-mini",
        "messages": [{
            "role": "user",
            "content": "Hello!"
        }]
    }'

{
  "id": "chatcmpl-xxxxxxx",
  "object": "chat.completion",
  "created": 1742738518,
  "model": "gpt-4o-mini-2024-07-18",
  "choices": [
    {
      "index": 0,
      "message": {
        "role": "assistant",
        "content": "Hello! How can I assist you today?",
        "refusal": null,
        "annotations": []
      },
      "logprobs": null,
      "finish_reason": "stop"
    }
  ],
  "usage": {
    "prompt_tokens": 9,
    "completion_tokens": 10,
    "total_tokens": 19,
    "prompt_tokens_details": {
      "cached_tokens": 0,
      "audio_tokens": 0
    },
    "completion_tokens_details": {
      "reasoning_tokens": 0,
      "audio_tokens": 0,
      "accepted_prediction_tokens": 0,
      "rejected_prediction_tokens": 0
    }
  },
  "service_tier": "default",
  "system_fingerprint": "xxxxx"
}

[nezhyborets]

Hi guys, @CamilleScholtz @sakeven , toolCalls and annotations are not optional, please let me know if it works for you

[CamilleScholtz]

Seems to work again, thanks!

[ndimitrakas]

Hi guys, @nezhyborets, I was curious if there was a way to make this work on 0.3.7, I gave up on it because I just found out about the update to 0.3.8. What should the structure look like if tools becomes non-optional again in the future? Below is a code snippet I was trying to implement for 0.3.7.

let tools: [ChatQuery.ChatCompletionToolParam] = [.init(function: .init(name: "is_profane", description: "Return a boolean value indicating whether the given string contains profanity", parameters: .init(type: .object, properties: ["inputText": .init(type: .string)], required: ["inputText"]))) ]

[nezhyborets]

@ndimitrakas hi, do you mean you wanted to fork 0.3.7? Or do you mean you wanted to make it work without changing the code for the SDK?

If fork, take a look at line 44 here. 25442c2. It controls how tool_calls is decoded, specifically it controls whether it's required for the field to be present in a response.

If without changing the code of SDK, there was no ability to work around the issue in 0.3.7. Middlewares functionality was merged just yesterday that enables an ability for a workaround in such case, and I believe we are going to add more options on how to be in cases when response does not much our structs that are responsible for parsing

[ndimitrakas]

@nezhyborets I meant without changing the code for the SDK, sorry for being unclear! But thank you for answering!!! Sounds great! 👍