Compile any video into deterministic, structured JSON and Markdown — giving your AI agents instant video understanding without burning tokens or wasting context window on heavy frame analysis.
No network calls. No API keys. No data leaves your machine. Whisper, PaddleOCR, and Qwen run entirely locally.
Treat video understanding as a compilation step. The exact same video and config always produces byte-identical facts.
Built-in FastAPI server. Give your AI agents the ability to submit videos and poll for JSON results via standard HTTP calls.
Use the Typer-based CLI for quick extraction in bash scripts, or import the typed Python SDK into your custom RAG pipelines.
Every extractor sits behind a shared interface. Swap Whisper for llama.cpp without forking or breaking the entire pipeline.
JSON artifacts carry a strict schema version validated by Pydantic. Feed it directly into LLMs without fear of silent drift.
Invoke the compiler from your code, get fully typed Pydantic models back. No complex async graph orchestration required.
from vindex import compile_video
# Compiles video and returns typed models
index = compile_video("lecture.mp4")
# Iterate through semantically grouped scenes
for scene in index.timeline.scenes:
print(f"Scene: {scene.id}")
print(f"Caption: {scene.caption}")
# Access all OCR text within this scene
for text in scene.ocr_events:
print(f"Found text: {text.text}")