Skip to content

web-slate/gitai

Repository files navigation

GitAI

GitAI is a project aimed at assisting developers in learning and using Git commands effectively. This README provides an in-depth look at the various approaches we've explored and their technical implementations.

1. Fine-tuned LLMs

Hugging Face Link

We fine-tuned two smaller language models to run locally:

  • Qwen2 0.5B
  • Gemma 2B

Training Data:

  • Git documentation
  • Stack Overflow Q&A tagged with git, github, and gitlab

Fine-tuning Process:

  • Used LORA (Low-Rank Adaptation) technique for fine-tuning
  • Preprocessed the training data to create input-output pairs
  • Fine-tuned the models on a MacBook Air M2 (16GB) with MLX-LM

Inference:

  • Models can be run locally on CPU or GPU
  • Input: Natural language query about Git
  • Output: Relevant Git command or explanation

2. AI Agent

Technologies Used:

  • LangChain Agents
  • ShellTool for executing terminal commands
  • OpenAI GPT-3.5 as the language model

How it works:

  1. User inputs a natural language request
  2. LangChain Agent processes the input using GPT-3.5
  3. Agent determines the appropriate Git command
  4. ShellTool executes the command in the terminal

Error Handling:

  • Uses language model as a reasoning engine to determine which actions to take and in which order
  • Agent can interpret error messages and take actions on its own

3. Commonly Used Commands Retrieval

Data Structure:

  • JSONLine file containing common Git commands, their descriptions, parameters, and syntax

Retrieval Process:

  1. User input is converted into an embedding using SentenceTransformers
  2. Chroma vector store performs cosine similarity search
  3. Top K most similar commands are retrieved

Technologies Used:

  • LangChain
  • Chroma as the vector store
  • SentenceTransformers for embedding

4. RAG QA Bot

Knowledge Base:

  • Git documentation
  • Stack Overflow Q&A tagged with git, github, and gitlab

Technologies Used:

  • EmbedChain for RAG
  • Mistral 7B as LLM (Hugging Face Inference)
  • Chroma as the vector store
  • SentenceTransformers for embedding

How it works:

  1. Knowledge base is chunked and embedded
  2. User question is embedded using the same model
  3. Searches in Vector store for relevant chunks
  4. Retrieved chunks are used to construct a prompt
  5. Language model generates an answer based on the prompt

Conclusion

After thorough testing and evaluation:

  1. The Commonly Used Commands Retrieval system proved to be the most efficient for quick, local operations. Its speed and accuracy make it ideal for beginners who want to learn Git.

  2. The AI Agent approach, while requiring an internet connection and OpenAI API, offers the most flexible and comprehensive solution, capable of handling complex queries and exectuing them on it's own.

Releases

No releases published

Packages

No packages published

Contributors 3

  •  
  •  
  •  

Languages