simple scritps for Integration of Autogen and memGPT to test differnt outcomes of LLM agents cooperation
This repository serves as a versatile playground for testing the cooperation and integration of different Large Language Models (LLMs) such as Autogen and memGPT. By leveraging Autogen and memGPT packages, users can explore various outcomes and interactions between these LLM agents within a controlled environment.
HOW TO USE:
Requirements:
An OpenAI API endpoint or clone(LMStudio works well, its the software used for the testing)
Autogen and memGPT packages installed locally.
https://microsoft.github.io/autogen/
https://memgpt.ai
For local LLM i recomed LMStudio for its simplicity and compatibility (specailly on amd hardware try out https://lmstudio.ai/rocm )
Clone this repository or copy the code in the examples to run.
Hook up your API in the scripts config
Modify the agents promts (using memgpt you should delete the agents if they will have different sistem promts and the same name since they are stored in a database when created)
Modify the task promt
run! and see if the hallucinate
results using 7B and 13B models with a token context size of 8192
natural-functions-Q8_0.gguf: Function Calling: Really good at function calling , few errors with memGPT Coding: can write simple scripts and single document python code, prone to hallucinations Story-Telling: prone to halucinate
llama-2-13b-chat.Q5_K_M.gguf Function Calling: Bad at function calling, cant be effectibly used with memGPT, works with Autogen alone Coding: Not good, can write REALLY simple code Story-Telling: Can produce good results if prompted with specifics about the story, loose track when memGPT kicks off managing context
dolphin-2.6-mistral-7b-dpo-laser.Q8_0.gguf Function Calling: The best model for memGPT, prone to errors formating autoGEN functions some times specially closing the SH tags for CLI execution Coding: Average, can be used to revise code with VScode plugins (Copilot clones like twinny) Story-Telling: prone to loose track and give really short statements instead of a story
airoboros-l2-13b-3.1.1.Q6_K.gguf Function Calling: works for memGPT, prone to forget autogen functions Coding: can write simple scripts , cant be tested further since it does not format correctly tags for autogen Stor-Telling: prone to loose trak, prone to respond empty (may be a VRAM limitation)
deepseek-coder-6.7B-instruct-GGUF Function Calling: not usable for memGPT, prone to forget autogen functions Coding: can write simple scripts the best, can modify documents, and kinda undesrtand the code, prone to halluciante nonsensical code when asked to resolve a complex task. Stor-Telling: incoherent
Resources: https://memgpt.readme.io/docs/autogen
Purpose:
This repository serves as a dynamic testing ground for experimenting with various LLM agents, aiming to understand their capabilities, limitations, and potential synergies. It facilitates exploration into how different models interact and perform across diverse tasks, offering insights into their behavior and efficacy in real-world applications. aimed to help anyone tryng to set it up for testing porpuses.