Description
There are a number of use cases where you'd want LLM tools calls to NOT be executed locally but instead returned to the client to trigger a client-side action. For example:
- dynamic visualizations/plotting (see this repository)
- updating UI components on the frontend
- implementing "callbacks" across service boundaries
Expected Behavior
I was thinking about adding a method to FunctionCallBack that looks like this:
public interface FunctionCallback {
/**
* @return whether this function is a client-executable function. Client-executable
* functions are not handled locally but returned to the client with the response.
*
*/
default boolean isClientFunction() {
return false;
}The isToolCall method in AbstractToolCallSupport would then check whether this method returns true for a particular tool call and return false if not. That would bypass the local tool execution and return the tool call as part of the response so that the user can then return it as needed to the frontend.
The rest of the tooling implementation could stay the same and the current work on extending function calling support would all carry over.
Current Behavior
I couldn't find a way to NOT invoke registered tools immediately. The only workaround I found to allowing tools to be passed to the frontend is to throw an exception in the function execution, raise a specific error with the payload, and then catch it. But that approach has a lot of downsides - not to mention very hacky ;-)
Context
See above. Let me know if there is a workaround. If not, I'm happy to contribute the described change or let me know if there is a better way.