feat: enable MCP to models without function calling ability #1978
Conversation
TommeyChang
changed the title
TommeyChang
changed the title
echo-yiyiyi
left a comment
echo-yiyiyi
left a comment
There was a problem hiding this comment.
A very cool MCPAgent! Left some comments.
Also, I’d like to know—could you briefly explain what model_fc_available means and why we need these two modes?
| ```json | ||
| { | ||
| "server_idx": idx, | ||
| "tool_name": "tool_name", | ||
| "tool_args": { | ||
| "arg1": value1, | ||
| "arg2": value2, | ||
| ... | ||
| } | ||
| } | ||
| ``` |
There was a problem hiding this comment.
Camel provides modules for structured output and scraping. Maybe you can check if it's available to avoid reinventing the wheel.
https://github.com/camel-ai/camel/blob/master/examples/structured_response/structure_response_prompt_engineering.py
https://github.com/camel-ai/camel/blob/master/examples/structured_response/json_format_response.py
There was a problem hiding this comment.
In the existing implementation, Deepseek-R1, gemmas and other models without function calling ability can not be used.
The model_fc_avaible is to set whether the large model has the function call ability.
I will change it into function_calling_available and try to use the modules for structured output and scraping.
| model: Optional[BaseModelBackend] = None, | ||
| model_fc_available: bool = False, | ||
| ) -> None: | ||
| r"""Initialize the MCP agent. |
There was a problem hiding this comment.
Maybe also add the reference of MCP Agent design here if there is any?
There was a problem hiding this comment.
thanks @TommeyChang ! As @yiyiyi0817 mentioned it would be great if we can use structured output feature instead of using prompt-based method, added one more PR here: feel free to help with review~
camel/agents/mcp_agent.py
Outdated
| await self._mcp_toolkit.disconnect() | ||
|
|
||
| def add_mcp_tools(self): | ||
| assert self._mcp_toolkit.is_connected(), "Server is not connected." |
There was a problem hiding this comment.
better not use assert in production code
There was a problem hiding this comment.
OK, I will fix it.
|
Check out this pull request on See visual diffs & provide feedback on Jupyter Notebooks. Powered by ReviewNB |
|
History messed up for review, moving to #2128. |
Pull request was closed
…ix for a clear git history) (#2128) Co-authored-by: tommey <tommeychang@163.com> Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com> Co-authored-by: Wendong <w3ndong.fan@gmail.com> Co-authored-by: Wendong-Fan <133094783+Wendong-Fan@users.noreply.github.com>
|
hey @TommeyChang , this feature has been merged in PR #2128, thanks so much for the contribution! Feel free to close this PR at your convenience |
|
close as feature merged in another PR |
Description
In this commit, a model without built-in function calling ability can use a MCP tools from giving the tool_name and tool_args.
In this case, we can set agents to equip with the same type of tools reducing the search space.
Checklist
Go over all the following points, and put an
xin all the boxes that apply.Fixes #issue-numberin the PR description (required)pyproject.tomlanduv lockIf you are unsure about any of these, don't hesitate to ask. We are here to help!