How to do Offline Recommendations with SQLite and Gemini

When working with a CMS (like PocketBase) it is common to add some sort of recommendatios for related content. For example you can have a list of blog posts, and show related posts either by random selection or recently viewed.

I first learned about this technique from Aaron Francis on his YouTube channel:

Text Embeddings

Text embeddings are a way to convert a chunk of text into a an array of numbers. Having a mathematical representation means we can easily store them in a database and run common functions to calculate the distances between vectors that we have stored.

You will need an API Key from AI Studio to generate the descriptions and embeddings.

In order to create the embedding we need to first generate chunk small enough to fit in the embedding window size. For example we can use an LLM like Gemini to generate a description for a blog post and then vectorize the description which we can store in the database.

We only need to generate a new embedding and description when the content changes which limits the billing costs to the frequency of the content changes.

Storing the Vectors