Learn how AI-powered search lets you find content inside videos. Scholaris indexes full content with semantic embeddings for meaning-based search.
The Problem with Traditional Video Search
Video content is traditionally unsearchable. You cannot Ctrl+F a lecture recording or a documentary. Finding a specific moment in a 90-minute video means scrubbing through the timeline manually, relying on vague chapter markers, or hoping someone added timestamps in the comments. For researchers and students working with recorded lectures, interviews, or video sources, this makes video content effectively inaccessible for reference purposes.
How AI Search Works
AI video search works through multiple layers. First, the audio track is extracted and transcribed using speech-to-text AI (Parakeet TDT), producing a full transcript with word-level timestamps. Then, keyframes are extracted and analyzed visually using multimodal embeddings (Qwen3-VL). Finally, both the transcript and visual content are embedded into a shared semantic space, allowing you to search by meaning across both what was said and what was shown.
Step-by-Step Workflow
1. **Upload your video** -- Scholaris accepts common video formats (MP4, MKV, WebM, etc.).
2. **Automatic transcription** -- The audio track is transcribed with word-level timestamps and speaker identification.
3. **Visual analysis** -- Keyframes are extracted and embedded for visual search.
4. **Search by text or concept** -- Type a query and find the exact moment in the video, whether it was spoken or shown visually.
5. **Jump to timestamp** -- Click any result to jump directly to that point in the video player.
6. **Cite with timestamps** -- Generate citations that include the exact timestamp for the referenced content.
Scholaris Capabilities
Scholaris makes video content as searchable as text. Key capabilities include:
- **Full transcription**: Automatic speech-to-text with word-level timestamps, powered by Parakeet TDT.
- **Speaker diarization**: Identify who is speaking at any given moment, critical for interviews and panel discussions.
- **Visual search**: Search for visual content shown in the video, not just spoken words.
- **Cross-modal queries**: Search with text and find matching video frames, or describe a scene to locate it.
- **Timestamp-linked citations**: Generate academic citations with precise timestamps.
Frequently Asked Questions
What video formats does Scholaris support?
Scholaris supports all common video formats including MP4, MKV, WebM, AVI, and MOV. The audio track is extracted using FFmpeg, which handles virtually any format.
Can Scholaris identify different speakers in a video?
Yes. Scholaris uses speaker diarization (pyannote) to identify and label different speakers. This is particularly useful for interviews, panel discussions, and multi-participant lectures.
How long does it take to process a video?
Processing time depends on your hardware. With a GPU, a 1-hour video typically takes 15-20 minutes. On CPU only, expect 1-2 hours for the same video.
Search inside any document with AI
Scholaris uses AI-powered semantic search to find answers across PDFs, videos, audio, and more — all running locally on your machine.