Skip to content

VideoIndex

VideoIndex turns long videos into a queryable knowledge base. Index a video once (decode, transcribe, read on-screen text, embed frames and text, detect shots and scenes) and ask questions that come back with the exact moment cited.

  • Rust core, sandboxed media decoding, ONNX models on CPU or GPU.
  • Python binding with operators and agent policies written in Python.
  • HTTP API, SSE and MCP server (vidx serve), so any agent framework can use an index as a tool.
  • Provider-agnostic: OpenAI-compatible servers, Anthropic, Gemini, local ONNX.

Try the demo over 36 hours of AI Engineer workshops and the Berkeley Agentic AI MOOC, or start with the quickstart.

The results pages compare VideoIndex with Gemini agentic video on questions that span a whole library and on the LVBench and MINERVA long-video benchmarks.

Questions, ideas, or something you are building on VideoIndex? Join the developers on Discord. Bugs and feature requests go to GitHub issues.