PromptGit (pgit / pvcs) 🚀

Version Control for LLM Prompts > Treat your prompts like code. Track history, measure costs, and catch regressions.

PromptGit (command: pvcs) is a local-first CLI tool designed to bring Git-like version control semantics to Prompt Engineering. It helps you manage prompt templates, track changes, and record execution metrics (latency, cost, tokens) in a local SQLite database.

✨ Features

Git-like Workflow: Familiar commands like init, add, commit, log, diff, checkout adapted for prompts.
Execution Tracking: Run prompts against LLM providers (Mock, OpenAI, Anthropic) and log every result.
Cost Analysis: Track exactly how much your prompt engineering is costing you per model.
Jinja2 Templating: Support for dynamic prompts using standard {{ variable }} syntax.
CI/CD Ready: Export analytics to JSON for integration with build pipelines.
Local-First: All data is stored in .pvcs/db.sqlite - you own your data completely.

📖 User Guide

1. Initialization

Start by creating a new folder for your prompt project and initializing pvcs.

mkdir my-prompts
cd my-prompts
pvcs init

2. Creating a Prompt Template

Create a text file using Jinja2 syntax for dynamic variables.

sentiment_analyzer.txt:

You are a helpful AI. Analyze the sentiment of the following text:
"{{ text }}"

Return JSON: {"sentiment": "positive|neutral|negative", "score": 0.0-1.0}

3. Tracking Changes

Add the file to the staging area and commit it.

pvcs add sentiment_analyzer.txt
pvcs commit -m "Initial draft of sentiment analyzer"

4. Running a Prompt & Analytics

Run prompts with data inputs and track the costs immediately.

pvcs run sentiment_analyzer.txt --data input.json --model gpt-4
pvcs analytics # Shows total runs, costs, and breakdown by model

🏗 Architecture

Repo: .pvcs/ directory contains db.sqlite and index.json.
Database: SQLite stores commits, tracked_files, and runs.
Content: Files are hashed (SHA-256) and stored in file_blobs for efficient deduplication.

🛠️ Tech Stack

Core: Python 3.12+
CLI Framework: Typer & Rich for a beautiful terminal experience.
Storage: SQLite for robust local data management.
Templating: Jinja2 for flexible prompt construction.

PromptGit (pvcs)