Local LLM Git Commit Summarizer

This project sets up two open-source LLMs locally Mixtral-8x7B-Instruct and Phi-3 Mini. It uses Ollama to run AI summarization tasks entirely offline.

It includes a Node.js script that:

• Retrieves commit messages from your Git repo
• Sends them to the selected model
• Returns a press-release-style summary for changelogs, tag messages, etc.

---

System Requirements

• macOS (Apple Silicon preferred)
• Ollama installed and running locally
• Node.js v18+ (for native fetch, fs/promises support)

---

Installation

1. Install Ollama

brew install ollama

Then start the Ollama server manually:

ollama serve &

Do not use brew services start ollama unless you want it running constantly.

Alternatively, there is a shell script included locally that you can source to add start / stop commands to your environment.

---

2. Pull the Models

Mixtral

ollama pull mixtral

Phi-3 Mini

ollama pull phi3:mini

---

3. Install Dependencies

This script uses only built-in Node.js modules, but make sure you have Node 18+ installed:

node -v

If needed:

brew install node

---

Usage

From inside any Git repo:

node summarize_commits.js [model]

Where [model] is one of:

• mixtral
• phi3:mini

If omitted, it will prompt you to choose interactively.

Alternatively, the included shell script -- assuming you sourced it earlier -- adds a summarize_commits command that can be run from any repo.

---

Example

node summarize_commits.js mixtral

Produces output like:

Introduced support for OAuth and fixed a session timeout bug affecting mobile users. Dependencies were also updated to improve stability.

---

What It Does

• Uses git describe and git log to get commit messages since the last tag
• Builds a natural-language prompt with that history
• Sends it to Ollama’s local model endpoint
• Streams and sanitizes the response
• Prints a one-paragraph customer-facing summary

---

Development Notes

• Uses raw http.request for low-dependency local calls
• Models are served on http://localhost:11434 by Ollama
• Responses are newline-delimited JSON chunks
• Any malformed chunks or control characters are stripped before parsing

---

Uninstall / Cleanup

To remove models:

ollama rm mixtral
ollama rm phi3:mini

To stop Ollama:

ollama stop all