Readiness and capacity: FFmpeg, yt-dlp and its JavaScript runtimes, the transcription and embedding backends, limits, and index size. ready is false, with reasons in problems, when ingest would fail.

Show what a link offers before downloading: title, length, and each quality with its size and whether it has sound on its own or needs the audio track merged in (most qualities above 360p do).

Save a link to a file. Fetches video and audio streams in parallel fragments and muxes them into one playable file, preferring H.264/AAC for MP4 so it opens everywhere. Returns a Job whose result is a DownloadReport.

Download (for a URL), transcribe, sample frames and embed, so the video can be searched and questioned. Speech and frame work run in parallel. Re-ingesting reuses finished stages, so a failed or interrupted job resumes. Returns a Job; its result holds the video_id, counts and timings.

Start one ingest job per source. Playlist and channel URLs expand to their first playlist_limit videos. Jobs run DVQA_MAX_CONCURRENT_JOBS at a time and the rest queue; a bad source is reported in skipped without stopping the others.

A job's status, stage and progress, optionally waiting for it to finish. Progress is also streamed as MCP progress notifications while waiting.

Recent download and ingest jobs, newest first.

Stop a queued or running job. Stages that already finished stay in the index and are reused.

The moments most relevant to a question, as evidence: for each, the transcript around it with timestamps and the frames from it as images (in moments order). Answer from the evidence and cite times. With verify, each moment carries calibrated support probabilities and evidence.supported says whether the video answers the question at all.

Ranked hits from three channels fused by weighted reciprocal rank: meaning-based transcript search, keyword (BM25) search, and text-to-image search over frames. 'The slide with the pricing table' finds the frame even if nobody said it. Each hit lists the channels that matched it, with rank and raw score.

Where a picture, logo, chart or slide appears: frames that look like it, and passages about it.

Every transcript segment containing the exact phrase, in time order. Use it to count mentions or collect every place a name or number comes up; use video_search for meaning rather than wording. A phrase split across two segments is not matched.

What was said, for the whole video or a range: timestamped lines, plain text, SRT, WebVTT, or JSON segments. Page through long videos with start and end.

Frames as images, at given timestamps or spread across a range. With exact=true each is decoded from the source at that instant and at higher resolution than the indexed samples, for reading small text.

Chapters with spans, opening words and distinguishing keywords. Uses the uploader's chapters when the source has them; otherwise finds topic shifts in the transcript embeddings (frames for silent video).

Cut a time range into its own MP4 with audio. Stream copy is instant; exact re-encodes.

Every video in the index with its id, title, length, status and index counts.

Source, dimensions, language, index counts, completed stages and the uploader's description.

Delete a video's transcript, frames and vectors. A file you ingested by path is never deleted; only the server's own downloaded copy and extracted frames are.