Build a Movie Recommendation Engine with pgvector

Recommendation systems are widely used in modern applications to suggest relevant content to users. They are commonly powered by vector embeddings and similarity search.

Instead of relying on a dedicated vector database, PostgreSQL® can perform vector similarity search using the pgvector extension.

In this tutorial, we deploy a simple movie recommendation system using PostgreSQL® with pgvector on Scalingo and a small Python API.

Planning your Deployment

• In this tutorial, we choose to deploy both the backend and frontend of the application on a single web container of size M.

• Performing similarity search requires a vector-capable database. In this tutorial, we choose Scalingo for PostgreSQL® addon with the pgvector extension enabled for this role.

• The application is written in Python, using Flask, FastAPI and LangChain to query the embeddings.

The recommendation workflow follows these steps:

1. Generate embeddings for each movie description.
2. Store these embeddings in the Scalingo for PostgreSQL® database, using pgvector.
3. Generate embeddings for user queries.
4. Retrieve the most similar movies using vector similarity search.

Deploying the Application

Using the Command Line

1. Clone our repository:

   git clone https://github.com/Scalingo/filmreco
   cd filmreco
   

2. Create the application on Scalingo:

   scalingo create my-filmreco
   

   Notice that our Command Line automatically detects the git repository, and adds a git remote to Scalingo:

   git remote -v
   
   origin   https://github.com/Scalingo/filmreco (fetch)
   origin   https://github.com/Scalingo/filmreco (push)
   scalingo git@ssh.osc-fr1.scalingo.com:my-filmreco.git (fetch)
   scalingo git@ssh.osc-fr1.scalingo.com:my-filmreco.git (push)
   

3. Provision a Scalingo for PostgreSQL® Starter 512 addon:

   scalingo --app my-filmreco addons-add postgresql postgresql-starter-512
   

4. Everything’s ready, deploy to Scalingo:

   git push scalingo main
   

Scalingo detects the Python environment, installs dependencies and starts the application. Your recommendation API is deployed.

Loading the Dataset

To populate the database with movies and embeddings, run the setup script as a one-off task:

   scalingo --app my-filmreco run python3 setup.py

The script enables the pgvector extension, creates the movies table and inserts the embeddings from the dataset to our database. The DATABASE_URL environment variable is automatically available in the one-off context.

For more information about running one-off tasks, refer to the Scalingo documentation on tasks.

Querying the Recommendation Engine

Open your deployed application on Scalingo and enter a prompt to retrieve the most similar movie using pgvector similarity search.

You now have a working recommendation system built with Python, pgvector, and LangChain on Scalingo. With this setup, you can start experimenting with more advanced use cases such as product recommendations, document similarity search, or other semantic search applications.