Embeddings are a good starting point for the AI curious app developer

🌈 Abstract

The article discusses the use of vector embeddings, a type of AI technology, in the context of an app developer building an icon discovery app. It covers the implementation details, including the choice of using Postgres with the pgvector extension, creating embeddings, implementing search and filtering, and providing similar recommendations. The article also provides a summary of the key decisions and sources for other options.

🙋 Q&A

[01] Embeddings and Postgres with pgvector

1. Why did the author choose to use Postgres with the pgvector extension instead of a specialized vector database? The author chose Postgres with pgvector because they didn't want to have to compare results across multiple data sources. Pgvector allows them to blend embedding search with business logic like filtering and scoring within the same database.

2. How does the author create the embeddings for the icons? The author creates a function that encodes the icon data, including the name, categories, and tags, into a string. This string is then used to generate the vector embeddings using an OpenAI embedding model.

3. What are the benefits of using vector embeddings for search compared to traditional text-based search? Vector embeddings allow for similarity-based search, where the results are ranked based on the cosine similarity between the search query and the icon embeddings. This enables more relevant results, even for queries in different languages, without the need for manually curated tags or a bag of words.

[02] Implementing Search and Recommendations

1. How does the author implement the search functionality using the vector embeddings? The author uses a SQL query to calculate the cosine similarity between the search query embedding and each icon embedding, orders the results by similarity, and limits the output to the top 50 results.

2. How does the author implement the similar recommendations feature? For icons that belong to a category, the author uses the same cosine similarity approach to find other icons in the same category. For icons without categories, the author compares the embedding of the selected icon to the embeddings of other icons in the same set to find similar recommendations.

3. What additional algorithm does the author mention to improve the search results over time? The author suggests an algorithm that combines the cosine similarity from the embedding search with the click data for each icon, giving more weight to icons that users have clicked on more often for a particular search query.

[03] Implementation Decisions and Resources

1. What are the different vector database options mentioned in the article besides Postgres with pgvector? The article mentions that there are many specialized vector databases to choose from, but the author chose Postgres with pgvector to blend embedding search with business logic.

2. What are the different client libraries and hosting options mentioned for working with pgvector? The author used Typescript and the drizzle-orm library to work with pgvector. The article also mentions using the elixir library with Ecto, as well as hosting options like Neon, fly.io, and Supabase.

3. What are the different embedding models and distance metrics discussed in the article? The author used OpenAI's text-embedding-3-small model and cosine similarity as the distance metric. The article also mentions that other embedding models and distance metrics like l2 distance and inner product are supported by pgvector.