Pixeltable

Tracked

A declarative data infrastructure for multimodal AI workloads that simplifies storage, indexing, and inference.

Author Pixeltable Open Sourced 2023-05-10 Last Commit Unknown

Pixeltable is an open-source declarative backend for multimodal AI applications that unifies the storage, indexing, transformation, and inference of images, video, audio, and documents under a single table interface. It replaces hand-built ETL scripts with incremental, versioned computations so teams can focus on model logic rather than pipeline plumbing.

Native Multimodal Types

  • First-class column types (pxt.Image, pxt.Video, pxt.Document) treating media alongside structured fields
  • Declarative computed columns that define transformation and inference pipelines once
  • Automatic incremental execution and caching to avoid redundant recomputation

Built-in Search & Retrieval

  • Embedding indexes and semantic search on any column without external vector infrastructure
  • Similarity retrieval and RAG workflows directly within the table abstraction
  • Supports retrieval-augmented generation, automated labeling, and object-detection pipelines

Extensibility & Integration

  • Custom UDFs and iterators for extending the system with domain-specific logic
  • Pre-built adapters connecting to OpenAI, Hugging Face, YOLOX, and other popular services
  • External media storage with PostgreSQL-managed metadata and view-maintenance for freshness
  • Apache-2.0 licensed with an active contributor community