Journal AI - Semantic Search using Vector Embeddings

A PostgreSQL + pgvector system to store journal entries and answer natural language questions via vector similarity search.

Setup Instructions

1. Install Dependencies

npm install

2. Set up PostgreSQL with pgvector

Run the Docker container with pgvector support:

docker run -d \
  --name journal-db \
  -p 5432:5432 \
  -e POSTGRES_PASSWORD=postgres \
  ankane/pgvector

3. Initialize Database Schema

Connect to the PostgreSQL database and run the schema:

psql -U postgres -h localhost -d postgres -f schema.sql

4. Configure Environment Variables

Edit .env and add your OpenAI API key:

OPENAI_API_KEY=sk-...

5. Start the Server

npm start

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

API Endpoints

Save a Journal Entry

POST /journal

Request:

{
  "entry": "played tennis with nicey after work"
}

Response:

{
  "status": "saved"
}

Ask a Question

POST /ask

Request:

{
  "question": "what racket sports did i play in the last 2 weeks?"
}

Response:

{
  "answer": "You played tennis, badminton, and squash in the last two weeks.",
  "matches": [
    {
      "id": 1,
      "entry": "played tennis with nicey after work",
      "created_at": "2026-03-08T06:34:01.517Z"
    },
    {
      "id": 2,
      "entry": "tried badminton with coworkers",
      "created_at": "2026-02-20T06:34:02.000Z"
    },
    {
      "id": 4,
      "entry": "played squash at the gym",
      "created_at": "2026-02-28T06:34:03.000Z"
    }
  ]
}

How It Works

1. Question is converted to an embedding using OpenAI's text-embedding-3-small model
2. PostgreSQL pgvector performs similarity search using the <-> operator (L2 distance)
3. Top 5 matching entries are retrieved from the database
4. GPT-4o-mini generates a natural language answer using the retrieved entries as context

Testing with Sample Data

# Seed the database
npm run seed

# Or add entries manually
curl -X POST http://localhost:3000/journal \
  -H "Content-Type: application/json" \
  -d '{"entry": "played tennis with sam after work"}'

# Ask a question
curl -X POST http://localhost:3000/ask \
  -H "Content-Type: application/json" \
  -d '{"question": "what racket sports did i play in the last 2 months?"}'