Skip to main content
Once uploaded, a video moves through a background pipeline before it’s searchable. See Upload Videos for the API calls that trigger and poll it, and Processing Status for the status values a job moves through.

​
Pipeline stages

  1. Probe — duration, resolution, fps, and codec are read from the file
  2. Transcription — the full audio track is transcribed by Whisper and written as timestamped speech segments
  3. Scene detection — shot boundaries are found via frame-difference analysis (CPU-only, runs for every plan). Cut points are added to the frame extraction grid so shots aren’t missed by the fixed interval alone
  4. Frame extraction — frames are pulled at regular intervals (every 5 seconds) plus at every detected cut, and scanned for CSAM before anything else touches them
  5. Visual embedding — each extracted frame is encoded into a vector (see Embeddings)
  6. Indexing — vectors and transcript segments are written to the vector store and Postgres, tagged with video_id and collection_id
  7. Scene labeling (Pay as you go and Enterprise plans) — a representative frame from each shot is classified into scene labels, and optionally captioned, then indexed as scene segments
  8. OCR — frames are sampled more densely (every 1.75s) and scanned for on-screen text, indexed as ocr segments
Once every stage completes, the job’s status flips to done and the video becomes searchable.
Scene labels and captions are Pay as you go/Enterprise features — a scene-type match in search results implies the source video was indexed on a paid plan. Frame, transcript, and OCR indexing run the same way on every plan.

​
Processing time

Typical processing time is 1–3 minutes per 10 minutes of video. Actual time depends on server load and video complexity — a video with long stretches of speech takes longer to transcribe than a silent one of the same length.

​
Partial failure handling

Frame extraction and embedding happen per-frame — if a handful of frames fail to decode, the pipeline skips them and continues rather than failing the entire job. Scene labeling and OCR are best-effort enrichment stages: if either fails outright, the job still completes with frame and transcript indexing intact.

​
Re-indexing

If Pureframe AI ships a new embedding model version, existing videos can be re-embedded without re-running transcription — frames already extracted and stored are simply re-encoded with the new model. This keeps re-indexing fast even for large libraries.