Prepare for the PgBouncer and IPv4 deprecations on 26th January 2024

Learn more

Build a content recommendation app with Flutter and OpenAI

2024-02-26

19 minute read

Recommending relevant content to the user is essential to keep the user interested in the app. Although it is a common feature that we would like to have in our apps, building it is not straightforward. This changed as vector databases and Open AI emerged. Today, we can perform semantic searches that are highly aware of the context of the content with just a single query into our vector database. In this article, we will go over how you can create a Flutter movie-viewing app that recommends another movie based on what the user is viewing.

A quick disclaimer, this article provides an overview of what you can build with a vector database, so it will not go into every detail of the implementation. You can find the full code base of the app in this article here to find more details.

Why use a vector database for recommending content

In machine learning, a process of converting a piece of content into a vector representation, called embeddings, is often used, because it allows us to analyze the semantic content mathematically. Assuming we have an engine that can create embeddings that are well aware of the context of the data, we can look at the distance between each embedding to see if the two content are similar or not. Open AI provides a well-trained model for converting text content into an embedding, so using it allows us to create a high-quality recommendation engine.

There are numerous choices for vector databases, but we will use Supabase as our vector database in this article, because we want to also store non-embedding data, and we want to be able to query them easily from our Flutter application.

What we will build

We will be building a movie listing app. Think Netflix except the users will not be able to actually view the movie. The purpose of this app is to demonstrate how to surface related content to keep the users engaged.

Application interfaces

Tools/ technologies used

  • Flutter - Used to create the interface of the app
  • Supabase - Used to store embeddings as well as other movie data in the database
  • Open AI API - Used to convert movie data into embeddings
  • TMDB API - A free API to get movie data

Creating the app

We first need to populate the database with some data about movies and its embeddings. For that, we will use the Supabase edge functions to call the TMDB API and the Open AI API to get the movie data and generate the embeddings. Once we have the data, we will store them in Supabase database, and query them from our Flutter application.

System overview

Step 1: Create the table

We will have one table for this project, and it is the films table. films table will store some basic information about each movie like title or release data, as well as embedding of each movie’s overview so that we can perform vector similarity search on each other.


_20
-- Enable pgvector extension
_20
create extension vector
_20
with
_20
schema extensions;
_20
_20
-- Create table
_20
create table public.films (
_20
id integer primary key,
_20
title text,
_20
overview text,
_20
release_date date,
_20
backdrop_path text,
_20
embedding vector(1536)
_20
);
_20
_20
-- Enable row level security
_20
alter table public.films enable row level security;
_20
_20
-- Create policy to allow anyone to read the films table
_20
create policy "Fils are public." on public.films for select using (true);

Step 2: Get movie data

Getting movie data is relatively straightforward. TMDB API provides an easy-to-use movies endpoint for querying information about movies while providing a wide range of filters to narrow down the query results.

We need a backend to securely call the API, and for that, we will use Supabase Edge Functions. Steps 2 through 4 will be constructing this edge function code, and the full code sample can be found here.

The following code will give us the top 20 most popular movies in a given year.


_32
const searchParams = new URLSearchParams()
_32
searchParams.set('sort_by', 'popularity.desc')
_32
searchParams.set('page', '1')
_32
searchParams.set('language', 'en-US')
_32
searchParams.set('primary_release_year', `${year}`)
_32
searchParams.set('include_adult', 'false')
_32
searchParams.set('include_video', 'false')
_32
searchParams.set('region', 'US')
_32
searchParams.set('watch_region', 'US')
_32
searchParams.set('with_original_language', 'en')
_32
_32
const tmdbResponse = await fetch(
_32
`https://api.themoviedb.org/3/discover/movie?${searchParams.toString()}`,
_32
{
_32
method: 'GET',
_32
headers: {
_32
'Content-Type': 'application/json',
_32
Authorization: `Bearer ${tmdbApiKey}`,
_32
},
_32
}
_32
)
_32
_32
const tmdbJson = await tmdbResponse.json()
_32
_32
const tmdbStatus = tmdbResponse.status
_32
if (!(200 <= tmdbStatus && tmdbStatus <= 299)) {
_32
return returnError({
_32
message: 'Error retrieving data from tmdb API',
_32
})
_32
}
_32
_32
const films = tmdbJson.results

Step 3: Generate embeddings

We can take the movie data from the previous step and generate embedding for each of them. Here, we are calling the Open AI Embeddings API to convert the overview of each movie into embeddings. overview contains the summary of each movie, and is a good source to create embedding representing each of the movies.


_20
const response = await fetch('https://api.openai.com/v1/embeddings', {
_20
method: 'POST',
_20
headers: {
_20
'Content-Type': 'application/json',
_20
Authorization: `Bearer ${openAiApiKey}`,
_20
},
_20
body: JSON.stringify({
_20
input: film.overview,
_20
model: 'text-embedding-3-small',
_20
}),
_20
})
_20
_20
const responseData = await response.json()
_20
if (responseData.error) {
_20
return returnError({
_20
message: `Error obtaining Open API embedding: ${responseData.error.message}`,
_20
})
_20
}
_20
_20
const embedding = responseData.data[0].embedding

Step 4: Store the data in the Supabase database

Once we have the movie data as well as embedding data, we are left with the task of storing them. We can call the upsert() function on the Supabase client to easily store the data.

Again, I omitted a lot of code here for simplicity, but you can find the full edge functions code of step 2 through step 4 here.


_20
// Code from Step 2
_20
// Get movie data and store them in `films` variable
_20
...
_20
_20
for(const film of films) {
_20
// Code from Step 3
_20
// Get the embedding and store it in `embeddings` variable
_20
_20
filmsWithEmbeddings.push({
_20
id: film.id,
_20
title: film.title,
_20
overview: film.overview,
_20
release_date: film.release_date,
_20
backdrop_path: film.backdrop_path,
_20
embedding,
_20
})
_20
}
_20
_20
// Store each movies as well as their embeddings into Supabase database
_20
const { error } = await supabase.from('films').upsert(filmsWithEmbeddings)

Step 5: Create a database function to query similar movies

In order to perform a vector similarity search using Supabase, we need to create a database function. This database function will take an embedding and a film_id as its argument. The embedding argument will be the embedding to search through the database for similar movies, and the film_id will be used to filter out the same movie that is being queried.

Additionally, we will set an HSNW index on the embedding column to run the queries efficiently even with large data sets.


_14
-- Set index on embedding column
_14
create index on films using hnsw (embedding vector_cosine_ops);
_14
_14
-- Create function to find related films
_14
create or replace function get_related_film(embedding vector(1536), film_id integer)
_14
returns setof films
_14
language sql
_14
as $$
_14
select *
_14
from films
_14
where id != film_id
_14
order by films.embedding <=> get_related_film.embedding
_14
limit 6;
_14
$$ security invoker;

Step 6: Create the Flutter interface

Now that we have the backend ready, all we need to do is create an interface to display and query the data from. Since the main focus of this article is to demonstrate similarity search using vectors, I will not go into all the details of the Flutter implementations, but you can find the full code base here.

Our app will have the following pages:

  • HomePage: entry point of the app, and displays a list of movies
  • DetailsPage: displays the details of a movie as well as its related movies

_10
lib/
_10
├── components/
_10
│ └── film_cell.dart # Component displaying a single movie.
_10
├── models/
_10
│ └── film.dart # A data model representing a single movie.
_10
├── pages/
_10
│ ├── details_page.dart # A page to display the details of a movie and other recommended movies.
_10
│ └── home_page.dart # A page to display a list of movies.
_10
└── main.dart

components/film_cell.dart is a shared component to display a tappable cell for the home and details page. models/film.dart contains the data model representing a single movie.

The two pages look like the following. The magic is happening at the bottom of the details page in the section labeled You might also like:. We are performing a vector similarity search to get a list of similar movies to the selected one using the database function we implemented earlier.

Application interfaces

The following is the code for the home page. It’s a simple ListView with a standard select query from our films table. Nothing special going on here.


_48
import 'package:filmsearch/components/film_cell.dart';
_48
import 'package:filmsearch/main.dart';
_48
import 'package:filmsearch/models/film.dart';
_48
_48
import 'package:flutter/material.dart';
_48
_48
class HomePage extends StatefulWidget {
_48
const HomePage({super.key});
_48
_48
@override
_48
State<HomePage> createState() => _HomePageState();
_48
}
_48
_48
class _HomePageState extends State<HomePage> {
_48
final filmsFuture = supabase
_48
.from('films')
_48
.select<List<Map<String, dynamic>>>()
_48
.withConverter<List<Film>>((data) => data.map(Film.fromJson).toList());
_48
_48
@override
_48
Widget build(BuildContext context) {
_48
return Scaffold(
_48
appBar: AppBar(
_48
title: const Text('Films'),
_48
),
_48
body: FutureBuilder(
_48
future: filmsFuture,
_48
builder: (context, snapshot) {
_48
if (snapshot.hasError) {
_48
return Center(
_48
child: Text(snapshot.error.toString()),
_48
);
_48
}
_48
if (!snapshot.hasData) {
_48
return const Center(child: CircularProgressIndicator());
_48
}
_48
final films = snapshot.data!;
_48
return ListView.builder(
_48
itemBuilder: (context, index) {
_48
final film = films[index];
_48
return FilmCell(film: film);
_48
},
_48
itemCount: films.length,
_48
);
_48
}),
_48
);
_48
}
_48
}

In the details page, we are calling the get_related_film database function created in step 5 to get the top 6 most related movies and display them.


_106
import 'package:filmsearch/components/film_cell.dart';
_106
import 'package:filmsearch/main.dart';
_106
import 'package:filmsearch/models/film.dart';
_106
import 'package:flutter/material.dart';
_106
import 'package:intl/intl.dart';
_106
_106
class DetailsPage extends StatefulWidget {
_106
const DetailsPage({super.key, required this.film});
_106
_106
final Film film;
_106
_106
@override
_106
State<DetailsPage> createState() => _DetailsPageState();
_106
}
_106
_106
class _DetailsPageState extends State<DetailsPage> {
_106
late final Future<List<Film>> relatedFilmsFuture;
_106
_106
@override
_106
void initState() {
_106
super.initState();
_106
_106
// Create a future that calls the get_related_film function to query
_106
// related movies.
_106
relatedFilmsFuture = supabase.rpc('get_related_film', params: {
_106
'embedding': widget.film.embedding,
_106
'film_id': widget.film.id,
_106
}).withConverter<List<Film>>((data) =>
_106
List<Map<String, dynamic>>.from(data).map(Film.fromJson).toList());
_106
}
_106
_106
@override
_106
Widget build(BuildContext context) {
_106
return Scaffold(
_106
appBar: AppBar(
_106
title: Text(widget.film.title),
_106
),
_106
body: ListView(
_106
children: [
_106
Hero(
_106
tag: widget.film.imageUrl,
_106
child: Image.network(widget.film.imageUrl),
_106
),
_106
Padding(
_106
padding: const EdgeInsets.all(8.0),
_106
child: Column(
_106
crossAxisAlignment: CrossAxisAlignment.stretch,
_106
children: [
_106
Text(
_106
DateFormat.yMMMd().format(widget.film.releaseDate),
_106
style: const TextStyle(color: Colors.grey),
_106
),
_106
const SizedBox(height: 8),
_106
Text(
_106
widget.film.overview,
_106
style: const TextStyle(fontSize: 16),
_106
),
_106
const SizedBox(height: 24),
_106
const Text(
_106
'You might also like:',
_106
style: TextStyle(
_106
fontSize: 16,
_106
fontWeight: FontWeight.bold,
_106
),
_106
),
_106
],
_106
),
_106
),
_106
// Display the list of related movies
_106
FutureBuilder<List<Film>>(
_106
future: relatedFilmsFuture,
_106
builder: (context, snapshot) {
_106
if (snapshot.hasError) {
_106
return Center(
_106
child: Text(snapshot.error.toString()),
_106
);
_106
}
_106
if (!snapshot.hasData) {
_106
return const Center(child: CircularProgressIndicator());
_106
}
_106
final films = snapshot.data!;
_106
return Wrap(
_106
children: films
_106
.map((film) => InkWell(
_106
onTap: () {
_106
Navigator.of(context).push(MaterialPageRoute(
_106
builder: (context) =>
_106
DetailsPage(film: film)));
_106
},
_106
child: FractionallySizedBox(
_106
widthFactor: 0.5,
_106
child: FilmCell(
_106
film: film,
_106
isHeroEnabled: false,
_106
fontSize: 16,
_106
),
_106
),
_106
))
_106
.toList(),
_106
);
_106
}),
_106
],
_106
),
_106
);
_106
}
_106
}

And that is it. We now have a functioning similarity recommendation system powered by Open AI built into our Flutter app. The context used today was movies, but you can easily image that the same concept can be applied to other types of content as well.

Afterthoughts

In this article, we looked at how we could take a single movie, and recommend a list of movies that are similar to the selected movie. This works well, but we only have a single sample to get the similarity from. What if we want to recommend a list of movies to watch based on say the past 10 movies that a user watched? There are multiple ways you could go about solving problems like this, and I hope reading through this article got your intellectual curiosity going to solve problems like this.

Resources

Share this article

Build in a weekend, scale to millions