Glue logic for LLM addons. Part of thesis work to conect User->ChatDev->ToT->MemGPT->LLM together in a atempt to improve maximum project size. ChatDev will be modified to be able to make use of MemGpt to improve LLM memory, and ToT to improve LLM reasoning ability. Different permutations of improvements MemGpt, ToT, and LLM versions gpt3.5-turbo-16k, gpt-4-0125-preview, or gpt-4-1106-preview for a total of 12 combinations. Twelve different configurations of ChatDev will undergo experimental testing on a set of projects to analyze the impact of different parameters on cost, project size, and development time.
Normal use case and common components of the original ChatDev, ToT, MemGpt calling to OpenAI for cloud LLM compute. ChatDev, ToT, and MemGpt are colectivly reffered to as modules in this text.
The module user interface and LLM API call is replaced with glue logic code to call the next module in the configuration, the UI of ChatDev is not replaced as the module is common to all configurations instead of creating a separate UI node.
Illustration of glue logic with the modified ToT expanded to show the replaced UI and LLM API.
Project ran out of time trying to integrate ToT and MemGpt into ChatDev, all of the code basses need integration testing to ensure behavior does not change during integration. Testing had to be stopped due to funding constraints after testing gpt3.5-turbo-16k on all projects, gpt3.5-turbo-16k was tested in a one shot configuration wit all requirements to create some data to compare the gathered data to.
- Clone this repository
git clone --recurse-submodules https://github.com/HeWhoThreadsLightly/LLMglue.git - Open the downloaded repository
cd LLMglue - Setup conda
conda create -n LLMglue python=3.10 - Activate the conda enviroment
conda activate LLMglue - Install dependencies
pip3 install -r requirements.txt