rig-integration example does not reply the response of the tool, only reply input args

[Akagi201]

Describe the bug

The rig-integration example only reply me the Tool name and Inputs args, did not reply me the result of the mcp server tool.

To Reproduce
Steps to reproduce the behavior:

1. config the deepseek_key and cohere_key in the config.toml
2. cargo run in the folder examples/rig-integration
3. Input commands like: "show me the git status of current directory"

Expected behavior

The bot should reply the MCP Tool used, Inputs argument and the missing tool call result.

Logs

The robot only reply Tool and Inputs, no tool call result

MCP Manager created, 1 servers started
get tool: git_status
get tool: git_diff_unstaged
get tool: git_diff_staged
get tool: git_diff
get tool: git_commit
get tool: git_add
get tool: git_reset
get tool: git_log
get tool: git_create_branch
get tool: git_checkout
get tool: git_show
Enter :q to quit
user> show me the git status of current directory
🤖 Agent: To check the git status of the current directory, I would need to use a tool that can execute git commands. Based on the instructions, I should provide the tool name and inputs rather than performing the operation directly.

Tool: git_command
Inputs: ["status"]
user> show me the git log of current directory
🤖 Agent: Tool: git_command
Inputs: ["log"]

[4t145]

I haven't find a way to carry the tool options on sending message when using rig with deepseek api. I basically copied the startup process from examples of rig but it doesn't work with deepseek.

[Akagi201]

After some debugging, I found that the CompletionModel's dynamic_tools only have one item, it's not populated correctly. So the input cannot match a proper tool. I will first try to use static_tools to bypass this in my project.

[jokemanfire]

I've written a simple-chat to use mcp tool , may it can help you it carries the information into the system prompt to meet the simplest requirements, but for more complex requirements, it is recommended to use a more comprehensive client that supports MCP, such as Claude

[jokemanfire]

Of course, thank you for fixing the example of rig

[Akagi201]

Now I can use static tools instead of dynamic tools for rig, and it works.

It's not the final solution, I may dig into it later.

[byeblack]

If we use the function calling of DeepSeek v3 (i.e. deepseek-chat), we only need to modify the example https://github.com/modelcontextprotocol/rust-sdk/blob/main/examples/rig-integration/src/mcp_adaptor.rs:

Add code:

impl rig::tool::ToolEmbeddingDyn for McpToolAdaptor {
 fn context(&self) -> serde_json::Result<serde_json::Value> {
 serde_json::to_value(self.tool.clone())
 }

 fn embedding_docs(&self) -> Vec<String> {
 vec![
 self.tool
 .description
 .as_deref()
 .unwrap_or_default()
 .to_string(),
 ]
 }
}

Then modify get_tool_set:

pub async fn get_tool_set(server: ServerSink) -> Result<ToolSet> {
 let tools = server.list_all_tools().await?;

 let mut tool_builder = ToolSet::builder();

 for tool in tools {
 eprintln!("get tool: {}", tool.name);
 let adaptor = McpToolAdaptor {
 tool: tool.clone(),
 server: server.clone(),
 };
 tool_builder = tool_builder.dynamic_tool(adaptor);
 }
 let tool_set = tool_builder.build();

 // let mut tool_set = ToolSet::default();
 // for tool in tools {
 // eprintln!("get tool: {}", tool.name);
 // let adaptor = McpToolAdaptor {
 // tool: tool.clone(),
 // server: server.clone(),
 // };
 // tool_set.add_tool(adaptor);
 // }

 Ok(tool_set)
}

Delete the context in main.rs to dynamically use the tools in MCP.

 let dpsk = openai_client
 .agent(deepseek::DEEPSEEK_CHAT)
// .context(
// r#"You are an assistant here to help the user to do some works. 
// When you need to use tools, you should select which tool is most appropriate to meet the user's requirement.
// Follow these instructions closely. 
// 1. Consider the user's request carefully and identify the core elements of the request.
// 2. Select which tool among those made available to you is appropriate given the context. 
// 3. This is very important: never perform the operation yourself and never give me the direct result. 
// Always respond with the name of the tool that should be used and the appropriate inputs
// in the following format:
// Tool: <tool name>
// Inputs: <list of inputs>"#,
// )
 .dynamic_tools(4, index, tool_set)
 .build();

Final Output(similar to the following result):

MCP Manager created, 1 servers started
get tool: git_status
get tool: git_diff_unstaged
get tool: git_diff_staged
get tool: git_diff
get tool: git_commit
get tool: git_add
get tool: git_reset
get tool: git_log
get tool: git_create_branch
get tool: git_checkout
get tool: git_show
store tool git_status
store tool git_add
store tool git_diff_unstaged
store tool git_commit
store tool git_diff_staged
store tool git_reset
store tool git_log
store tool git_show
store tool git_checkout
store tool git_create_branch
store tool git_diff
Enter :q to quit
user> show me the git status of current directory
🤖 Agent: args: {"repo_path":"./"}

🛠 Tool Call: {"content":[{"type":"text","text":"Repository status:\nOn branch main\n\nNo commits yet\n\nUntracked files:\n (use \"git add <file>...\" to include in what will be committed)\n\t.gitignore\n\tCargo.lock\n\tCargo.toml\n\tc\n\tconfig.toml\n\tsrc/\n\nnothing added to commit but untracked files present (use \"git add\" to track)"}],"isError":false}
🤖 Agent: 
user> show me the git log of current directory
🤖 Agent: args: {"repo_path":"./"}

🛠 Tool Call: {"content":[{"type":"text","text":"Reference at 'refs/heads/main' does not exist"}],"isError":true}
🤖 Agent: 
user> 

[byeblack]

This is almost certainly a problem with the example. In rig, the add_tool method of ToolSet can only add ToolType::Simple, and the schemas method can only export ToolType::Embedding.

[4t145]

@byeblack Nice try, could you fix this example please?