Full support for tools/functions

[carlos-verdes]

When using a tool that respond with a JSON (for example an API call) the agent is simply returning the JSON.

Doing some research I found out we need to pass the output of the tool response back to the model so it can build a proper human readable answer.

• OpenAI function calling
• Anthropic tool use

To support this we need some steps:

• update Message abstraction in completion module to add new field call tool_calls
• when the response of a model is to call a tool we need to get the output of the tool call + the initial tool call message (that's why previous point is needed) and send again the prompt + these messages to the model
• we should support also calls to more than one function (I see some implementations using tool_calls.first() which is not correct).

Example of new Message:

#[derive(Clone, Debug, Deserialize, Serialize)]
pub struct Message {
    /// "system", "user", or "assistant"
    pub role: String,
    pub content: String,
    #[serde(default)]
    pub tool_calls: Vec<ToolCall>,
}

#[derive(Clone, Debug, Deserialize, Serialize)]
pub struct ToolCall {
    pub call_id: String,
    pub function_name: String,
    pub function_params: serde_json::Value,
}

Then we can add a logic similar to this in completion implementation (one per provider):

impl completion::CompletionModel for CompletionModel {
    type Response = CompletionResponse;

    #[cfg_attr(feature = "worker", worker::send)]
    async fn completion(
        &self,
        mut completion_request: CompletionRequest,
    ) -> Result<completion::CompletionResponse<CompletionResponse>, CompletionError> {
...
                   match response.choice {
                        completion::ModelChoice::ToolCall(name, id, args) => {
                            let tool_call_message = ...;
                            completion_request.chat_history.push(tool_call_message);

                           let tool_call_response = ....;
                           completion_request.chat_history.push(tool_call_response);

                          // recursive call with updated context
                          self.completion(completion_request)
...
}

I started a new branch in my fork to cover these points and a first tentative for OpenAI.

Graphical representation of where we are and what is missing:

[carlos-verdes]

@cvauclair take a look into this please

[0xMochan]

Hey, thanks for the post! In the future, please make an issue for related requests like this as that is the proper channel for us to catalog changes and improvements to the rig project, thanks!

Check out #199 for active progress on this!

[carlos-verdes]

This may imply more than one issue that's why I used a discussion.
The first issue is already covered by one of your PRs which is to improve Message model

[carlos-verdes]

Opened a PR for this (still in development)

#276