Enhancement: Optimize the logic of agents' tool invocation.

[Passerby1011]

What features would you like to see added?

Currently, the ability of agents to invoke tools entirely depends on the model's inherent Function Calling capability. Some models are exceptionally proficient, yet they may lack Function Calling capabilities or perform poorly in this aspect, rendering them unsuitable for use in agents, such as o1.

Can the tool invocation and the main model of agents be separated, allowing models without Function Calling capabilities to be effectively applied in agents as well?

At the same time, the current use of agents entirely relies on the model's capabilities itself; LibreChat has not made any interventions in this regard. This approach overly depends on the model's inherent abilities. Clearly, with the current method, the user experience of agents has only reached a basic, entry-level usability, far from being considered user-friendly. I believe this is an area in urgent need of improvement, even though it is challenging.

The system of LibreChat coordinates the use of various models and tools to achieve better results, rather than relying solely on the capabilities of the models themselves.

I believe that for a long time to come, the capabilities of large models will hardly see a qualitative leap, and even if there were, the average person would find it difficult to afford their exorbitant prices.

More details

notiing

Which components are impacted by your request?

No response

Pictures

No response

Code of Conduct

• I agree to follow this project's Code of Conduct

[danny-avila]

Can the tool invocation and the main model of agents be separated

The definition of an agent requires the LLM to be agentic. I think what you're asking for is for some kind of chain-of-thought process, chaining agents together.

I have already implemented this and might just add it to the main repo, however it's very experiemental.

This way o1 can do a lot of the "thinking" and a smaller tool-calling model like gpt-4o-mini can call the tools

[Passerby1011]

Yes, the current model's capabilities are not yet sufficient for us to fully rely on it.

[Passerby1011]

We need additional systems to guide the model in completing tasks.

[starascendin]

that would be sick