[Bug] vllm docking dialogue with MCP will fail

[zhainanshidai]

Contact Information

No response

MaxKB Version

1.10.8

Problem Description

vllm对接含MCP的对话会对接失败，提示 -

Steps to Reproduce

1.构建vllm环境，确保环境正常

2.构建高级编排
只有对话一个节点，提问showtables 不带mcp

3.加入mcp

提问

vllm报错（此处可以看出，是请求参数出错了）：

换其他应用，使用百炼平台的千问：

换成本地vllm：

使用dify +mcp+vllm：正常提问

查阅资料发现，vllm目前仅支持OpenAI规范API，加入mcp时，会导致出错。这里需要增强下对vllm的兼容性，对有mcp的情况做处理

The expected correct result

No response

Related log output

Additional Information

No response

[liqiang-fit2cloud]

MCP工具的调用效果取决于LLM的质量和MCP服务的质量，建议你尝试更换大模型、计算平台或者MCP。
MaxKB本身没有也不会对某个大模型或某个计算平台做特殊处理，个别模型返回数据的差异会由ollma、vllm等内部处理，它们暴露接口的返回数据不会有变化。

[shaohuzhang1]

Bot detected the issue body's language is not English, translate it automatically. 👯👭🏻🧑‍🤝‍🧑👫🧑🏿‍🤝‍🧑🏻👩🏾‍🤝‍👨🏿👬🏿

---

The calling effect of MCP tools depends on the quality of LLM and the quality of MCP services. It is recommended that you try to replace the large model, computing platform or MCP.
MaxKB itself does not and will not do special processing for a large model or a computing platform. The differences in data returned by individual models will be processed internally by ollma, vllm, etc., and the data returned by their exposed interfaces will not change.

[liqiang-fit2cloud]

MaxKB用的客户端是https://github.com/langchain-ai/langchain-mcp-adapters
这块没有别的逻辑，就几行代码，不会对某个平台某个模型做特殊处理。
ollama、vllm这类本地运行的推理服务对于各种模型的支持有很多问题，你可以找找有没有你遇到的问题，https://github.com/ollama/ollama/issues?q=is%3Aissue%20qwen3
https://github.com/vllm-project/vllm/issues?q=is%3Aissue%20mcp
肯定不如阿里百炼上的qwen3，建议有条件还是用在线的。

[shaohuzhang1]

Bot detected the issue body's language is not English, translate it automatically. 👯👭🏻🧑‍🤝‍🧑👫🧑🏿‍🤝‍🧑🏻👩🏾‍🤝‍👨🏿👬🏿

---

The client used by MaxKB is https://github.com/langchain-ai/langchain-mcp-adapters
There is no other logic in this piece, just a few lines of code, and it will not deal with special treatments on a certain platform or a model.
Locally run inference services such as ollama and vllm have many problems with the support of various models. You can find out if there are any problems you encounter. https://github.com/ollama/ollama/issues?q=is%3Aissue%20qwen3
It is definitely not as good as qwen3 on Alibaba Bailian. It is recommended to use online if possible.

[dengpei89]

这个问题似乎跟maxkb有点关系，我用cherry studio测试这个qwen3-14b（vllm方式部署得）大模型调用mcp就一切正常。而maxkb里勾选mcp就报错，不选mcp就正常。