A major compute cost in Figma's AI-search indexing pipeline came from finding designs inside files and rendering their thumbnails, according to its October 2024 engineering account. Generating the numerical representations used for similarity search was not what drove the main compute costs.
That finding directed optimization toward preparing the searchable material. Figma rewrote its file-parsing logic, moved thumbnail rendering to cheaper processors, and found from its data that refreshing the index at most once every four hours would mean processing only 12% of the data. The saving came with a clear product choice: search updates could lag behind edits.
Terms in this piece · Glossary
embedding — A list of numbers representing a piece of text's meaning, so that similar meanings end up numerically close and can be searched.
