Tool Call Repair

ToolCallRepairMiddleware is a middleware designed to automatically repair invalid tool calls from large language models.

When large language models output tool call parameters that conform to a JSON Schema, they may sometimes generate malformed JSON content (often in the arguments field) due to model limitations. Such failed parsing attempts are flagged by LangChain and stored in the invalid_tool_calls field. ToolCallRepairMiddleware automatically detects this field and uses the json-repair library to attempt to fix the formatting, allowing the tool calls to execute normally.

Usage Notes

Before using this middleware, ensure langchain-dev-utils[standard] is installed. Refer to the Installation Guide for details.

Parameter Description

This middleware is designed for zero-configuration out-of-the-box use. No parameters are required during instantiation.

Usage Examples

Standard Usage

from langchain_dev_utils.agents.middleware import ToolCallRepairMiddleware

agent = create_agent(
    model="vllm:qwen2.5-7b",
    tools=[run_python_code, get_current_time],
    middleware=[
        ToolCallRepairMiddleware()
    ],
)

Convenient Usage (Recommended)

Since instantiating ToolCallRepairMiddleware requires no configuration parameters, this library provides a pre-configured global instance tool_call_repair. It is recommended to use this directly to simplify your code:

from langchain_dev_utils.agents.middleware import tool_call_repair

agent = create_agent(
    model="vllm:qwen2.5-7b",
    tools=[run_python_code, get_current_time],
    middleware=[tool_call_repair],
)

Important Notes

This middleware cannot guarantee 100% repair of all invalid tool calls. The actual repair effectiveness depends on the capabilities of the json-repair library. Furthermore, it only operates on the invalid calls within the invalid_tool_calls field.