The above post is nonsense. The AI generated slop doesn’t even show the tool call.
If the AI has a tool call open, you have to match IDs in the immediate tool response.
messages: user → function call → tool return : the AI generates in response to the return.
On Chat Completions, and not on Assistants, and not on Responses with internal tools, you can call the API with the tool return in either of these patterns of image injection:
messages: user → [user image] → function call → tool return : the AI thinks the user added the image as prompt for it to call a function.
or
messages: user → function call → tool return → [user image] : the AI doesn’t directly respond to the tool result, it thinks the user added an image.
Neither is optimal, are not how the models were trained on functions, and a different quality AI will break the understanding further. A tool return array role message that can take multiple parts for images (just like the user message) is required.
OpenAI is unable to fulfill this articulated need, apparent ever since DevDay 2023 introduced “tool” and “vision”, but is off writing nonsense.