pixeltable

Why we included this project

Most teams building multimodal AI apps end up stitching together object storage, a vector index, an orchestrator, and a few ad-hoc scripts, each with its own quirks to babysit. Pixeltable collapses that into a single Python system: you define tables with native types for video, images, audio, and documents, then attach computed columns that trigger API calls or local models the moment new rows arrive. Chunking, frame extraction, embedding indexing, and agent tool calls become plain database operations instead of separate pipelines you maintain by hand. Everything is versioned, with transactions, caching, retries, and observability built in, so you can query an old snapshot or roll back a change. For data engineers and ML folks who want training features, RAG corpora, and serving data to live in one queryable, reproducible store, that's a real difference from scattered artifacts.