Build a 'more like this' recommender with embeddings

Recommenders come in two families. Collaborative filtering learns from behaviour: it watches who liked what and finds patterns across people, so it can tell you "readers who bought this also bought that" without knowing a single thing about the books themselves. It is what powers most big recommendation systems, and it is powerful, but it has a cold opening problem: with no ratings yet, it has nothing to learn from.

Content-based filtering works the other way. It looks at the item itself, its title, its description, its text, and recommends other items whose content is similar. It needs no ratings and no other users at all: one item is enough to find things like it. That is what you build here. It is the right tool when you are starting out, when your catalogue changes often, or when you want recommendations to run entirely on your own machine with nothing uploaded.

The engine behind it is the same one from the semantic search lesson: a small model turns each item into a list of numbers that stands for its meaning, and "similar" becomes "close together". Download the kit into a new folder: save recommend.py and requirements.txt in the folder itself, and films.json inside a sub-folder called catalogue. Then set it up and see the headline result:

cd $HOME\Documents
mkdir recommender; cd recommender
# put recommend.py and requirements.txt here, and films.json in a catalogue\ sub-folder
uv venv --python 3.13
uv pip install -r requirements.txt
uv run recommend.py build
uv run recommend.py similar "The Matrix"

cd ~/Documents
mkdir recommender && cd recommender
# put recommend.py and requirements.txt here, and films.json in a catalogue/ sub-folder
uv venv --python 3.13
uv pip install -r requirements.txt
uv run recommend.py build
uv run recommend.py similar "The Matrix"

build reads the 80 films and turns each into numbers (the first run downloads the 90MB model, so give it a minute). similar then finds the films most like The Matrix:

More like "The Matrix" (1999) [sci-fi]
  1. 0.674  The Matrix Reloaded (2003) [sci-fi]
  2. 0.492  Ex Machina (2014) [sci-fi]
  3. 0.416  The Dark Knight (2008) [action]
  4. 0.382  Casino Royale (2006) [action]
  5. 0.381  Blade Runner 2049 (2017) [sci-fi]

The closest film is its own sequel at 0.674, then more science fiction and action. Here is the part worth holding on to: the model was never told the genre. It read only each film's title and one-line description and worked out that these belong together. The [sci-fi] and [action] labels are printed for you to judge the results; the recommender never sees them. The rest of the lesson takes this apart and shows you how to recommend from a person's taste, not one film.