Datie AI - Modern Dating Matchmaker

A powerful AI-driven matchmaker for dating apps that uses Gemini embeddings and Pinecone vector database to find highly compatible matches based on personality, interests, and lifestyle. It uses Google's Gemini AI to understand the "vibe" of a user's profile (their personality, interests, and lifestyle) by converting their profile into a numerical representation called an "embedding." These embeddings are stored in a Pinecone vector database, which allows for very fast and efficient searching of similar users.

It also ultilize Rule-based Compatibility Filtering which applies a set of rules to filter and rank potential matches based on user preferences like age, gender, relationship goals, and other "deal-breakers."

Features

• AI-Powered Matching: Uses Google's Gemini API to generate semantic embeddings from user profiles
• Vector Similarity Search: Leverages Pinecone for fast and accurate similarity matching
• Smart Compatibility Filtering: Filters matches based on preferences (age, gender, lifestyle, education)
• Comprehensive Scoring: Combines embedding similarity (60%) and preference compatibility (40%)
• RESTful API: Simple endpoints to embed users and find matches

How It Works

1. Profile to Vector Conversion

Each user profile (bio, interests, lifestyle, occupation) is converted into a rich text representation, then transformed into a 768-dimensional vector using Gemini's text-embedding-004 model.

2. Vector Storage

All user vectors are stored in Pinecone, a specialized vector database optimized for similarity search.

3. Similarity-Based Matching

When finding matches for a user:

1. Generate an embedding for the user's profile
2. Query Pinecone to find the nearest neighbor vectors (cosine similarity)
3. Filter results based on compatibility preferences
4. Calculate a combined score: 60% similarity + 40% compatibility
5. Return the top matches sorted by overall score

Prerequisites

• Node.js (v18+)
• Gemini API Key - Get one here
• Pinecone Account - Sign up here

Setup

1. Clone and Install

npm install

2. Configure Environment Variables

Copy .env.example to .env and fill in your credentials:

cp .env.example .env

GEMINI_API_KEY=your_gemini_api_key_here
PINECONE_API_KEY=your_pinecone_api_key_here
PINECONE_INDEX_NAME=datie-ai-matchmaking

3. Create Pinecone Index

In your Pinecone dashboard, create a new index with:

• Name: datie-ai-matchmaking (or your chosen name)
• Dimensions: 768
• Metric: cosine

4. Start the Server

npm run dev

The server will start on http://localhost:3000

API Endpoints

GET /

Get API information and available endpoints

Response:

{
  "message": "Datie AI - Modern Dating Matchmaking Engine",
  "endpoints": {
    "POST /embed-users": "Embed all user profiles and store in Pinecone",
    "GET /match/:userId": "Find matches for a specific user",
    "GET /users": "Get all users",
    "GET /users/:userId": "Get a specific user"
  }
}

GET /users

List all available users

Response:

{
  "total": 50,
  "users": [
    {
      "userId": "dc2a5a7b-e91b-4b1a-9c7f-3d6e5a4b1c2d",
      "firstName": "Ethan",
      "age": 29,
      "gender": "Male",
      "location": { "city": "Boulder", "state": "CO" },
      "bio": "Environmental scientist who feels most at home..."
    }
  ]
}

GET /users/:userId

Get details for a specific user

Example: GET /users/dc2a5a7b-e91b-4b1a-9c7f-3d6e5a4b1c2d

POST /embed-users

Important: Run this once to initialize the vector database

Embeds all user profiles and stores them in Pinecone.

Response:

{
  "message": "Embedding complete",
  "total": 50,
  "success": 50,
  "failed": 0,
  "errors": []
}

GET /match/:userId?topK=10

Find matches for a specific user

Parameters:

• userId (path): The user ID to find matches for
• topK (query, optional): Number of matches to return (default: 10)

Example: GET /match/dc2a5a7b-e91b-4b1a-9c7f-3d6e5a4b1c2d?topK=5

Response:

{
  "message": "Matches found!",
  "forUser": {
    "userId": "dc2a5a7b-e91b-4b1a-9c7f-3d6e5a4b1c2d",
    "firstName": "Ethan",
    "age": 29,
    "gender": "Male",
    "location": { "city": "Boulder", "state": "CO" }
  },
  "totalMatches": 5,
  "matches": [
    {
      "userId": "c4d5e6f7-a8b9-4c1d-a2b3-d4e5f6a7b8c9",
      "firstName": "Sophia",
      "age": 30,
      "gender": "Female",
      "location": { "city": "San Diego", "state": "CA" },
      "bio": "Yoga instructor and wellness advocate...",
      "occupation": "Yoga Instructor",
      "interests": ["Yoga", "Meditation", "Beach", "Dogs"],
      "profileImages": ["https://example.com/..."],
      "similarityScore": 0.92,
      "compatibilityScore": 85,
      "overallScore": 89
    }
  ]
}

Testing the System

Step 1: Embed Users

First, embed all users into Pinecone:

curl -X POST http://localhost:3000/embed-users

This will process all users and store their embeddings in Pinecone (takes ~5-10 seconds).

Step 2: Find Matches

Get matches for a user (e.g., Ethan):

curl http://localhost:3000/match/dc2a5a7b-e91b-4b1a-9c7f-3d6e5a4b1c2d?topK=5

You'll get the top 5 most compatible matches!

Step 3: Explore Different Users

List all users to find different userIds:

curl http://localhost:3000/users

Then try matching for different personalities and preferences.

Matching Algorithm

The matching system uses a two-stage approach:

Stage 1: AI Similarity (60% weight)

• Converts profiles to semantic embeddings using Gemini
• Finds nearest neighbors using cosine similarity in Pinecone
• Captures personality "vibe" and lifestyle compatibility

Stage 2: Preference Compatibility (40% weight)

• Age Range: Must fall within both users' preferences
• Gender Preference: Must match what each is looking for
• Smoking/Drinking: Respects deal-breakers
• Children: Aligns family goals
• Education: Matches education preferences
• Pets: Respects "must love dogs/cats" preferences
• Shared Interests: Bonus points for common hobbies

Final Score

Overall Score = (Similarity × 0.6 × 100) + (Compatibility × 0.4)

Project Structure

datie-ai/
├── src/
│   ├── data/
│   │   └── users.json              # Sample user profiles
│   ├── services/
│   │   ├── embedding.ts            # Gemini embedding service
│   │   ├── pinecone.ts             # Pinecone client setup
│   │   └── matchmaking.ts          # Core matching logic
│   ├── utils/
│   │   ├── profileToText.ts        # Profile to text conversion
│   │   └── compatibility.ts        # Compatibility checking
│   └── index.ts                    # Main API server
├── .env.example                    # Environment variables template
├── package.json
├── tsconfig.json
└── README.md

Key Functions

profileToText(user: User): string

Converts a user profile into rich text that captures their personality, interests, and preferences.

getEmbedding(text: string): Promise<number[]>

Generates a 768-dimensional embedding vector using Gemini's text-embedding-004 model.

embedUser(user: User): Promise<void>

Embeds a single user and stores in Pinecone with metadata.

getCompatibilityScore(user1: User, user2: User): number

Calculates a 0-100 compatibility score based on preferences and shared interests.

findMatches(userId: string, allUsers: User[], topK: number): Promise<Match[]>

Main matching function that combines AI similarity with preference compatibility.

🌍 Production Considerations

1. Rate Limiting: Add rate limiting to embedding endpoints
2. Caching: Cache embeddings to avoid re-computing
3. Batch Processing: Process embeddings in background jobs
4. Monitoring: Add logging and error tracking
5. Security: Add authentication and authorization
6. Pagination: Implement pagination for large result sets
7. Distance Filtering: Add geographic distance filtering
8. Real-time Updates: Webhook for profile updates

📚 Technologies Used

• Hono: Fast, lightweight web framework
• Google Gemini: Advanced AI embeddings
• Pinecone: Vector database for similarity search
• TypeScript: Type-safe development
• Node.js: Runtime environment