How our recipe extraction pipeline works

Every time you import a recipe, we run it through four extraction stages, each faster and more reliable than the last.

1. Page-by-page: we use font-based title detection to find recipes without calling any model.
2. Cluster-based: if that fails, we look for dense ingredient groupings anywhere in the document.
3. Font-based: a simpler heuristic fallback.
4. Vision fallback: as a last resort, we render each page as an image and let a vision model read it directly.

The whole pipeline runs in parallel, caches aggressively, and falls back gracefully. Next time you import a popular cookbook, chances are we've seen it before and return results in a few hundred milliseconds.