Video RAG

Pipeline Overview

video.mp4
   |-- ffmpeg demux --------> audio.wav  --> Whisper --> transcript (timestamped segments)
   |                                                       |
   |                                                       v
   |                                         speaker-aware chunks
   |
   |-- PySceneDetect -------> scene list --> ffmpeg extract keyframes --> images
                                                                            |
                                                                            v
                                              CLIP / SigLIP / VoyageAI multimodal embeddings

Index:  text chunks (+ time) + image chunks (+ time) in vector store with shared schema.
Query:  embed text -> search text index AND multimodal index -> merge -> answer with timestamp.