RAG on Git repositories

This project is a solution to utilize Retrieval Augmented Generation on Git Repositories.

Requirements

Python >= 3.10

Windows

Windows requires additional Visual C++ 2015 build tools to be installed.
They are available on Microsoft website Build Tools

Selected models

Following model was selected as Default models for Embedding.

• Embedding Model: all-mpnet-base-v1
• serve-model project is needed for standalone model serving.

To change it, rewrite the default value in MODEL_TYPES.py. For example to embedding-3-small

The Default LLM model was seleted as:

• LLM Model: Llama-3.2-3B
• serve-model

But this is changable in the Config section, allowing user to choose OpenAI LLM of their choice.

Number of Retrieved Files

The approximate number of files retrived durring Document Retrival is defined in CONTEXT_SIZE.py.

• GIT_DOCUMENTS defines approximate number of documents that will be taken from all branches of git repo together
• GIT_DIVERSE_K defines the divese field from with most disimilar documents will be taken
• SHARED_DOCUMENTS defines approximate number of documents that will be taken from all requested zips
• SHARED_DIVERSE_K is same as GIT_DIVERSE_K

These numbers can be altered. The selected values were observed to be good midleground in information correctness and size.

How to Run

To run, clone this repository and install dependancies in following steps

pip install  -r requirements.txt

To run the solution, run python on script gradio_app_git_update.py

python gradio_app_git_update.py

This will produce a link to a frontend UI.
To utilize OpenAI models it is required to provide API key in the Config section.

Provide Companny API key

If maneged behined authetication system, API key can be set as a default option.
The user inputed API key will be used if inputed.

To set company wide key, set eviromental variable OPENAI_API_KEY that will be utilized by the application.

# Windows
$env:OPENAI_API_KEY=Company API key

# Linux
export OPENAI_API_KEY=Company API key

Custom Models

If custom models are to be utilized, follow the serve-model project and run custom Huggingface models. Custom models can be inserted to MODEL_TYPES.py with proper base_url where the model is served from.