PRSense

Production-Ready AI-Powered Pull Request Duplicate Detection

Mission: Build trustworthy tools that give repositories memory — reducing duplicate work while preserving contributor credit.

Automatically detect duplicate PRs and save hours of review time. Deploy in 5 minutes.

[!TIP] New Here? Start with START_HERE.md for a beginner-friendly overview, then check the guides below.

Quick Links:

• Installation: docs/install.md
• User Guide: docs/user-guide.md
• CLI Usage: docs/cli-usage.md
• Deployment: docs/deployment.md

Key Features

Core Capabilities

• 95% Accuracy — Real OpenAI embeddings (or free local alternatives)
• 2ms Latency — Bloom filter + ANN for sub-linear scaling
• Production Storage — SQLite (dev) or PostgreSQL + pgvector (prod)
• GitHub Bot — Auto-comments on duplicate PRs
• Deploy in 5min — Vercel, Docker, or manual server
• Cost-Effective — $0-50/month depending on scale
• Attribution Tracking — Preserves original authorship credit

Advanced Features

• Persistent Storage — Exports for SQLite/Postgres
• Score Breakdown — Explainable "Why" for every match
• Batch API — Check 100s of PRs in CI/CD pipelines
• Embedding Cache — Reduces API costs by 90%
• Configurable Weights — Tune scoring (Text vs Diff vs Files)
• Dry-Run Mode — specific testing without indexing
• ONNX Local Embeddings — 100% offline, privacy-first
• Cross-Repo Detection — Find duplicates across your entire org

📖 See Features Documentation for details.

Architecture

┌─────────────┐
│  New PR     │
└──────┬──────┘
       │
       ▼
┌─────────────────┐
│ Bloom Filter    │ ──── Early rejection
│ (Fast Path)     │
└──────┬──────────┘
       │ might contain
       ▼
┌─────────────────┐
│ Embedding       │ ──── Text + Diff embeddings
│ Pipeline        │
└──────┬──────────┘
       │
       ▼
┌─────────────────┐
│ ANN Search      │ ──── Candidate retrieval (k=20)
│ (Retriever)     │
└──────┬──────────┘
       │
       ▼
┌─────────────────┐
│ Multi-Signal    │ ──── Text + Diff + File Overlap
│ Ranker          │      (Configurable weights)
└──────┬──────────┘
       │
       ▼
┌─────────────────┐
│ Decision        │ ──── HIGH (≥0.9), MEDIUM (≥0.82), LOW
│ Engine          │
└──────┬──────────┘

Project Structure

prsenses-labs/
├── package.json                # Monorepo Root
├── PRSenses/                   # Core Library & API
│   ├── src/                    # Detection Logic
│   └── action/                 # GitHub Action
├── prsense-vscode/             # VS Code Extension
└── prsense-analytics/          # Analytics Dashboard

Quick Start

Installation

git clone https://github.com/prsense-labs/prsense
cd prsense
npm install     # Installs dependencies
npm run build   # Builds all packages

[!TIP] Zero setup mode: After building, the CLI works immediately using local ONNX embeddings. No API key required.

6 Ways to Use

1. GitHub Action (CI/CD)

- uses: prsense-labs/prsense@v1
  with:
    github-token: ${{ secrets.GITHUB_TOKEN }}
    embedding-provider: 'onnx'

2. VS Code Extension (Local Dev)

• Detects duplicates while you type.
• Extension Guide

3. CLI Tool (Manual Check)

npm run cli check my-pr.json

4. GitHub Bot (Webhook)

• Deploy to Vercel in 2 minutes.
• Deployment Guide

5. Library (NodeJS Integration)

import { PRSenseDetector } from 'prsense'
const result = await detector.check(prData)

6. Microservice (Docker/API)

• Run as a standalone REST API.
• Production Setup

Usage Example

Simple API (Recommended)

import { PRSenseDetector, createPostgresStorage } from './prsense.js'
import { createOpenAIEmbedder } from './embedders/openai.js'

// 1. Initialize
const storage = await createPostgresStorage().init()
const detector = new PRSenseDetector({
  embedder: createOpenAIEmbedder(),
  storage,
  enableCache: true
})

// 2. Check for duplicates
const result = await detector.checkDetailed({
  prId: 123,
  title: 'Fix login bug',
  description: 'Handle empty passwords',
  files: ['auth/login.ts']
})

// 3. Handle Result
if (result.type === 'DUPLICATE') {
  console.log(`Duplicate of PR #${result.originalPr}`)
  console.log(`Confidence: ${result.confidence}`)
  console.log(`Breakdown:`, result.breakdown)
  // Breakdown: { textSimilarity: 0.95, diffSimilarity: 0.88, ... }
}

Documentation Index

• problem.md — Problem statement
• algorithms.md — How it works (Deep Dive)
• data_structures.md — Bloom filters & Vector Index
• scoring.md — The math behind the match
• complexity.md — Performance analysis
• features.md — Detailed feature guide
• ROADMAP.md — Future plans

Design Principles

1. Scalability — O(1) lookups with Bloom filters
2. Accuracy — Multi-signal scoring > Single vector search
3. Privacy — Local embeddings (ONNX) support
4. Modularity — Swap storage, embedders, or rankers easily

What We Don't Do

PRSense tools are designed to assist humans, not replace them.

We do not:

• Block or auto-close pull requests
• Judge code quality
• Replace maintainer decisions or code review

Contributing

Contributions welcome! Please read CONTRIBUTING.md before submitting a PR.

• 💬 Questions? Start a Discussion
• 🐛 Found a bug? Open an Issue
• 📋 What's new? See the CHANGELOG

License

MIT License

PRSense — Because duplicate PRs waste everyone's time. Let's fix that.