Where data is stored
Video files and information about those videos are stored separately. When investigating a problem, first identify which kind of data you need to inspect.
From video to search data
videos stores the title, owner, file reference, transcript, processing state, and related metadata. The file itself is not stored in a DB row.
scene_embeddings enables semantic search of transcript segments. The API converts questions into numeric search vectors and finds similar scenes within permitted videos.
Questions and answers
chat_logs.response stores segments and server-owned sources. retrieved_contexts retains deduplicated scene text and course metadata. Partial stream events are not stored as separate chat records.
The stream's done follows answer persistence. Responses without a course are not saved. Retired evaluation tables retain historical data; no new scores are written.
Other storage locations
| Data | Storage or management |
|---|---|
| Browser login state | Better Auth session and browser cookies |
| External API keys saved by users | Encrypted in the database |
| MCP API keys and OAuth | apikey, oauth_*, and related tables |
| Undelivered jobs | external_tasks |
| Worker execution records | job_executions |
| Transcription and visual-cache checkpoints | video_transcription_* and video_visual_* |
| Visual caches for uploads | Private media/<file_key>.frames-v1.json and .focus-v1.bin objects in R2 / Garage |
Related: Data dictionary, ER diagrams, Job delivery and recovery.