fix: gpt-4o: fixed failing tool/function calling

[adubovik]

Usually, when the assistant replies with a tool call, the content is set to null.

This should be accounted for in the follow-up tokenization of the chat completion request. The following method doesn't do it:

ai-dial-adapter-openai/aidial_adapter_openai/gpt4_multi_modal/transformation.py

Lines 121 to 123 in fe3faea

	content = message.get("content", "")
	custom_content = message.pop("custom_content", {})
	attachments = custom_content.get("attachments", [])

It was made more forgiving:

        content = message.get("content") or ""
        custom_content = message.pop("custom_content", None) or {}
        attachments = custom_content.get("attachments") or []

How to reproduce the bug

Set the .env file:

GPT4O_DEPLOYMENTS=gpt-4o-2024-05-13

Send the following request to the adapter-openai running locally at port 5006:

MODEL=gpt-4o-2024-05-13
ENDPOINT=...
KEY=...

URL=http://0.0.0.0:5006/openai/deployments/${MODEL}/chat/completions?api-version=2024-02-01

curl -X POST $URL \
  -H "X-UPSTREAM-ENDPOINT:https://${ENDPOINT}.openai.azure.com/openai/deployments/${MODEL}/chat/completions" \
  -H "X-UPSTREAM-KEY:${KEY}" \
  -d '{
"stream": false,
"messages": [
  {
    "role": "user",
    "content": "What is the weather like in Boston and London today?",
    "custom_content": null
  },
  {
    "role": "assistant",
    "content": null,
    "custom_content": {
      "attachments": null
    },
    "tool_calls": [
      {
        "id": "call_Q6aBknWA6Jx9H5hvCEHpfJsv",
        "type": "function",
        "function": {
          "name": "get_current_weather",
          "arguments": "{\"location\": \"Boston, MA\", \"unit\": \"celsius\"}"
        }
      },
      {
        "id": "call_LNAJMTe5BddYZ49Te8QmHo9O",
        "type": "function",
        "function": {
          "name": "get_current_weather",
          "arguments": "{\"location\": \"London, UK\", \"unit\": \"celsius\"}"
        }
      }
    ]
  },
  {
    "role": "tool",
    "tool_call_id": "call_Q6aBknWA6Jx9H5hvCEHpfJsv",
    "content": "30"
  },
  {
    "role": "tool",
    "tool_call_id": "call_LNAJMTe5BddYZ49Te8QmHo9O",
    "content": "20"
  }
],
"tools": [
  {
    "type": "function",
    "function": {
      "name": "get_current_weather",
      "description": "Get the current weather in a given location",
      "parameters": {
        "type": "object",
        "properties": {
          "location": {
            "type": "string",
            "description": "The city and state, e.g. San Francisco, CA"
          },
          "unit": {
            "type": "string",
            "enum": ["celsius", "fahrenheit"]
          }
        },
        "required": ["unit", "location"]
      }
    }
  }
],
"tool_choice": "auto",
"temperature": 0.0
}'

It fails with the error:

  File "./aidial_adapter_openai/gpt4_multi_modal/transformation.py", line 152, in <listcomp>
    await self.transform_message(message) for message in messages
    ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "./aidial_adapter_openai/gpt4_multi_modal/transformation.py", line 128, in transform_message
    content_meta = await self.download_content_images(content)
                   ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "./aidial_adapter_openai/gpt4_multi_modal/transformation.py", line 102, in download_content_images
    for content_part in content:
TypeError: 'NoneType' object is not iterable

[adubovik]

/deploy-review

Environment URL: https://chat-ai-dial-adapter-openai-pr-175.nightly-test.deltixhub.io
E2E tests status: success