How it works
Find the teaching, demo, or explanation moment in your own video using visual and transcript search.
Video Moment Finder indexes sampled frames and available speech transcripts. Select a video from your private library and describe what you showed or said: text search can return both visual moments and spoken passages with timestamps. You can also upload an example image to find visually similar frames.
The pipeline
1. Upload or import
You upload a video file up to 90 minutes, or import a YouTube video you own. The upload is direct-to-storage via a presigned URL, so the file never round-trips through our API servers.
2. Frames and speech
A background worker samples video frames and transcribes speech in direct uploads. YouTube imports use available subtitle or automatic caption tracks; they do not fall back to speech transcription when captions are missing. Frames and transcript segments keep their source timestamps.
3. Multimodal embeddings (Qwen3-VL)
Qwen3-VL represents sampled frames as vectors for visual search. Available transcript segments are also embedded for semantic search, so a text query can retrieve a spoken explanation as well as something shown on screen.
4. Vector store
Visual and transcript vectors are stored for retrieval. Before searching, the API checks that you can access the selected video, then restricts retrieval to that video.
5. Search
A text query searches visual moments and available transcripts, returning timestamped visual and spoken matches. Spoken matches include transcript text; visual matches include thumbnails. An example-image query searches visual frames only. Review the source moment to check whether it answers your question.
Why Qwen3-VL
Qwen3-VL is an open-weight vision-language model trained jointly on images and text, which means a photo of a whiteboard and the phrase “a whiteboard with equations” end up close together in the same embedding space. That joint space is what makes text-to-frame and image-to-frame search work with the same index.
We picked it over alternatives for three reasons. First, it is open-weight, which matters for a project shipped under AGPLv3 and removes a hard dependency on a proprietary embedding API. Second, its training explicitly includes visual reasoning and on-screen text, both of which matter for real-world video (slides, UI recordings, signage). Third, the model is small enough to run cost-effectively at the scale of hundreds to thousands of frames per video without making the unit economics fall apart.
Accuracy tradeoffs
Best at
Finding a teaching, demo, or explanation moment in a lesson, webinar, or workshop you own. Try "a whiteboard diagram of a neural network" for a visual moment or "where I explain the difference between training and inference" for a spoken explanation.
Weaker at
Speech retrieval depends on the available transcript: transcription errors, unclear audio, or missing YouTube captions can hide relevant passages. Visually similar frames and details between sampled frames can also be missed. Results are candidates, not guaranteed exact matches.
Practical tips
Describe what you showed or explained, then review both visual and spoken results. If the first query misses, try a more specific topic or rephrase it. For image search, choose a clear example of the scene you want; an image query does not search spoken words.
Results are AI-generated and won't always be right — see the support page for more on expected accuracy and refunds.
Building on top
The same pipeline is exposed as an HTTP API and as a remote MCP server, so agents and scripts can upload, check status, and search videos directly. See the developers page for the MCP connector and OpenAPI reference.
Last updated 2026-10-26.