[BFCL] Function-calling parser returns empty response when tool_calls=[]

Author: ankush13rCreated Apr 27, 2026Updated Apr 27, 2026

Describe the issue _parse_query_response_FC does not correctly handle API responses where message.tool_calls exists but is an empty list ([]). In that case, the current implementation returns model_responses=[] instead of falling back to message.content.

ID datapoint

Datapoint / Model Handler permalink:
Issue: Function-calling parser returns empty model_responses when tool_calls=[]
Gorilla repo commit #: 6ea57973c7a6097fd7c5915698c54c17c5b1b6c8

What is the issue Some APIs return tool_calls=[] even when the model response is a normal text response available in message.content.

The current code only falls back to message.content inside the except block. However, when tool_calls is present as an empty list, no exception is raised (vllm returns empty tool_calls):

python

model_responses = [
    {func_call.function.name: func_call.function.arguments}
    for func_call in api_response.choices[0].message.tool_calls
]

This produces:

python

model_responses = []
tool_call_ids = []

As a result, the parser returns an empty response instead of the actual model content.

Proposed Changes

Add an explicit fallback when model_responses is empty after parsing tool_calls. Also keep the existing try/except behavior for legacy compatibility.

python

def _parse_query_response_FC(self, api_response: Any) -> dict:
    try:
        message = api_response.choices[0].message

        model_responses = [
            {func_call.function.name: func_call.function.arguments}
            for func_call in message.tool_calls
        ]
        tool_call_ids = [
            func_call.id for func_call in message.tool_calls
        ]

        if not model_responses:
            model_responses = message.content
            tool_call_ids = []

    except Exception:
        message = api_response.choices[0].message
        model_responses = message.content
        tool_call_ids = []

    model_responses_message_for_chat_history = message

    return {
        "model_responses": model_responses,
        "model_responses_message_for_chat_history": model_responses_message_for_chat_history,
        "tool_call_ids": tool_call_ids,
        "input_token": api_response.usage.prompt_tokens,
        "output_token": api_response.usage.completion_tokens,
    }

python

{
    "previous_datapoint": [],
    "updated_datapoint": [
        {
            "change": "Add fallback to message.content when parsed tool_calls produce an empty model_responses list.",
            "reason": "Some APIs return tool_calls=[] for non-function-call responses, which currently causes the parser to return an empty response instead of the textual content."
        },
        {
            "change": "Keep try/except fallback for legacy response shapes.",
            "reason": "Maintains backward compatibility with APIs that may not expose tool_calls or may use a different response schema."
        }
    ]
}

Additional context This issue affects function-calling model handlers when the API returns a valid assistant message with content, but also includes tool_calls=[]. The parser should treat an empty tool_calls list the same as no function call and return the text content instead.

Source: ShishirPatil/gorilla

View original on GitHub View discussion on GitHub