Xiaomi MiMo API: Fixing Tool Call Format Errors
If you're working with the Xiaomi MiMo API, specifically the mimo-v2-flash model, and you've encountered a perplexing 400 Bad Request: "Param Incorrect" - text is not present error when using tool or function calls, you're not alone. This issue arises after the model has successfully identified a tool to call and is attempting to proceed with the conversation. It's a common stumbling block that can halt your development progress, especially if you're expecting compatibility with established standards like OpenAI's function calling specification. This article aims to demystify this error, explain why it happens, and guide you toward a resolution so you can get back to building powerful AI applications with Xiaomi MiMo.
Understanding the Xiaomi MiMo Tool Call Error
The core of the problem lies in how the Xiaomi MiMo API handles messages that include tool_calls. In the widely adopted OpenAI specification, when an assistant's response contains tool_calls (meaning the AI has decided to use a specific function or tool), the content field within that message is permitted to be null. This design allows the API to clearly distinguish between a conversational response (which would have text content) and an instruction to execute a tool (where the primary information is in the tool_calls structure). However, the mimo-v2-flash model, as observed, deviates from this standard. It mandates that the text field must be present and non-empty, even when tool_calls are being utilized. This strict requirement, while seemingly a minor detail, creates a compatibility issue for developers relying on the OpenAI spec. It essentially forces a format that the model isn't designed to accept in that specific context, leading directly to the "Param Incorrect" - text is not present error. This can be particularly frustrating because the initial tool identification by the model might be perfectly correct, but the subsequent response formatting breaks the chain.
Why the text Field Matters (to MiMo)
Let's delve deeper into why Xiaomi MiMo seems to be insistent on the text field. While OpenAI's specification allows for content: null when tool_calls are present, it's crucial to remember that different models and APIs can have their own interpretations or specific implementation details. In the case of mimo-v2-flash, it appears the underlying architecture or the pre-processing layer expects a string value for the text field in every assistant message, regardless of whether it's a direct conversational reply or an instruction to execute a tool. This might be a design choice to simplify parsing on their end or a legacy requirement from earlier versions. The error message, "Param Incorrect" - text is not present, is quite literal: the API is checking for the existence and validity of the text field and failing when it's missing. This suggests that even if the tool_calls array is populated correctly with function names, arguments, and IDs, the absence of accompanying text triggers the validation error. Therefore, to successfully integrate with the Xiaomi MiMo API for tool calling, you need to adapt your output formatting to include a text field, even if it's just an empty string or a placeholder message, whenever tool_calls are being used.
Reproducing the Issue: A Step-by-Step Guide
To effectively troubleshoot and resolve the Xiaomi MiMo tool call format error, it's essential to understand precisely how to trigger it. The process typically involves a sequence of interactions between your application and the API. First, you initiate a conversation by sending a user message to the mimo-v2-flash model. This message must be crafted in a way that prompts the AI to identify and decide to use one of the tools or functions you've defined. This could involve asking a question that requires external data retrieval or an action that can only be performed by a specific tool. Second, upon receiving the model's response, which should ideally contain a tool_calls object detailing the function to be invoked and its arguments, you then need to package this information and send it back to the API. This is the crucial step where you're essentially informing the model about the tool it should execute. This might involve constructing a new API call where the messages array includes the original user query, the model's tool_calls response, and then your application's simulated response indicating the tool has been called. Third, after sending this
