AI Video Search Engine

AI Video Search Engine (AVSE) is a retrieval-augmented search demo for finding answers inside videos. The author built it to show that a search engine for video content, comparable to what Google does for text, is feasible with readily available tools, given how much knowledge now sits in short-form video.

When a video is stored, its transcription is extracted from YouTube along with thumbnail, description and title; the content is turned into vector embeddings with JigsawStack's multimodal embedding model, the transcript is chunked with timestamps, and everything is stored in Supabase PostgreSQL using pgvector. A search runs a cosine similarity query across the database, then a second search finds related chunks within each result video.

The stack combines Supabase for the database and auth, Hasura for the GraphQL layer hosted on Fly, JigsawStack for summaries, prompts and embeddings, and Vercel for the Next.js app. No license is listed in the repository. A hosted demo is available, and an older version using a smaller English-only embedding model is kept in an archive branch.