Pixeltable

Detailed Introduction

Pixeltable is an open-source declarative data infrastructure for multimodal AI applications. It exposes a unified table interface to manage images, video, audio, and documents, making ingestion, transformation, indexing and retrieval first-class capabilities. Pixeltable integrates with Large Language Models (LLM) and external Vector DBs to enable Retrieval-Augmented Generation (RAG) workflows while providing versioning and reproducibility for production workloads.

Main Features

• Unified multimodal table types: pxt.Image, pxt.Video, pxt.Document to manage diverse media.
• Declarative computed columns: define processing and model inference once; execution is incremental and cached.
• Built-in embedding indexes and semantic search: add embedding indexes directly on tables for similarity search and RAG.
• Broad model and service integrations: adapters for OpenAI, Hugging Face, YOLOX, and more.

Use Cases

• Multimodal retrieval and Q&A systems (RAG + LLM).
• Automated image/video labeling and object detection pipelines.
• Reproducible data pipelines that combine ETL, feature engineering, and model inference for production deployments.

Technical Features

• Incremental computation and view maintenance to reduce recomputation costs.
• Extensible UDFs and custom iterators for user-defined processing.
• Local caching and persistent metadata (Postgres) with media stored outside the DB.
• Apache-2.0 licensed with an active contributor community and sample apps.