Purpose
This local project organizes long-video material so a natural-language query can return an interval from the original media, its speech or visual basis, and an exportable clip for review.
Structure
The processing structure contains media ingestion, logical chunks, speech transcription, visual observations, a shared timeline, evidence retrieval, a derived document index, and clip export. Original media and analysis artifacts remain separate. Chunk, transcript, visual-observation, timeline, and index-document records are stored as JSONL files.
How it works
Each logical chunk stores its offset in the original video, and local timestamps are mapped to global time before storage. Queries read transcript, visual-observation, and timeline JSONL files, then combine their matching documents while preserving evidence references and export ranges. Index rebuilding gathers documents from those three sources and writes a manifest with total and per-source counts. An export bundle may contain video, audio, subtitles, representative frames, evidence data, and a manifest.
Current state
The project remains a private experiment. The MVP covers logical chunking, global time mapping, replaceable ASR and VLM backends, JSONL artifact loading, index rebuilding, evidence queries, and clip export. Artifact version fields, content-digest validation, invalid-source rejection, and global sorting during rebuild remain possible future hardening rather than current behavior.