> For a complete page index of the Captain API documentation, fetch https://docs.captain.dev/llms.txt?excludeSpec=true # Video > How Captain indexes video files. Clips of up to 120 seconds, each with a transcript and a written description of what is on screen, so a query lands on a time offset. ## Agent Quick Reference - Video * Extensions: `.mp4`, `.mov`, `.avi`, `.mkv`, `.webm`, `.flv`, `.wmv`. Normalised to a standard 720p MP4 before processing. * Segmented into contiguous clips of up to 120 s (fixed windows). Each clip gets one native embedding from the clip itself, a transcript slice, and a written description: category line, then subjects, actions, setting, on-screen text transcribed verbatim (slides, captions, UI), charts and diagrams. The description never paraphrases speech; that is the transcript's job. * Transcript: one speech-to-text pass over the whole file, language detected automatically; force it with `transcription_language` (for example `es-US`, `pt-BR`). Word timings are used to slice the transcript per clip. No speaker labels. * Query hit: `modality: "video"`, content `[Video: name]` + description + `Transcript: …`; time range at `location.start_seconds` / `location.end_seconds` (v3) or `metadata.startSec` / `endSec` (v2). * A file whose audio track cannot be decoded is indexed video-only; speech is still captured from the source file. * YouTube: `POST /v2/collections/{c}/index/youtube` with `mode` `transcript` (default), `audio` or `video`, `on_missing_transcript`, `languages`, up to 20 URLs. * Billing: 20 credits per minute of video. Upload limits by route on File Size Limits. * `mask_pii`: video is not masked. A request with `mask_pii: true` skips video files: they are not indexed, and the job lists them as skipped. Captain indexes video as clips of up to 120 seconds. Each clip has a time range, a native embedding of the clip, a transcript slice and a written description of what is on screen. This page covers how a video is cut, what each clip contains, what a result contains, YouTube options, and the limits. ## This is useful for * Product demos and webinars, where the reader wants the moment rather than the file * Training and onboarding video, found by what is on the slide as well as what is said * Screen recordings of bugs and walkthroughs, found by the text in the UI * Site and inspection footage filmed on a phone, with or without narration ## What one video becomes The file is normalised to a standard 720p MP4 and cut into contiguous clips of up to 120 seconds. Each clip gets a native embedding of the clip itself, a transcript slice and a written description, stored together on one record. ## How a video is chunked #### Clips with a time range Clips are contiguous 120-second windows; the last one is shorter. Every clip carries `location.start_seconds` and `location.end_seconds` on the v3 query, or `metadata.startSec` and `metadata.endSec` on v2. #### Transcript Speech-to-text runs once over the whole file, so a sentence that crosses a clip boundary is transcribed correctly. Word timings are then used to give each clip the words spoken inside its window. Language is detected automatically; set `transcription_language` (for example `es-US` or `pt-BR`) to force one. There are no speaker labels. #### Description Each clip's description has a category line and three to five sentences covering subjects, actions, objects and setting, with on-screen text transcribed verbatim (slide titles, captions, UI labels) and charts or diagrams read out. The description does not repeat the speech; spoken words are in the transcript. #### Three signals on one record The native embedding, the transcript slice and the description are stored on the same record. A query matches the picture by vector similarity and the words by text. The result text is `[Video: file name]`, then the description, then `Transcript: …`. #### Audio problems A video whose audio track cannot be decoded is indexed video-only. Its speech is still transcribed, because the speech-to-text pass reads the source file rather than the normalised copy. A silent video has no transcript. ## YouTube `POST /v2/collections/{collection}/index/youtube` accepts up to 20 URLs. `mode: "transcript"` (the default) indexes the captions as text and bills as text. `mode: "audio"` or `mode: "video"` runs the media pipeline and bills per minute. `on_missing_transcript` (`fail`, `audio` or `video`) sets what happens to a video with no captions, and `languages` sets the caption language preference. ## Limits and billing Video is billed at 20 credits per minute, summed per job. Size limits depend on the route: 100 MB through the direct upload endpoint, 500 MB from the dashboard, and no upload limit from a cloud storage connector. Collections that contain media rerank by default; see [Multimodal Search](/guides/multimodal-search) and [File Size Limits](/guides/file-size-limits). ## Examples #### Jump to a moment in a product demo Goal: return the clip where the presenter filters a report by region, with the offset to play from. **`request`** ```json title="request" { "query": "filtering the report by region", "limit": 3 } ``` **`result (one item)`** ```json title="result (one item)" { "modality": "video", "text": "[Video: q3-demo.mp4]\nCategory: software demo, dashboard. A presenter opens a filter panel on the left of an analytics dashboard and selects 'EMEA'...\nTranscript: let's narrow this down to EMEA so you can see the regional split", "location": { "start_seconds": 480.0, "end_seconds": 600.0 } } ``` The description records the filter panel and the transcript records "EMEA". Both are on the clip that starts at 480 seconds. #### Training video, found by the slide title Goal: find the clip that shows a procedure's slide, even though the presenter does not read the title aloud. **`request`** ```json title="request" { "query": "lockout tagout procedure", "limit": 3, "filter": { "course": { "$eq": "site-induction" } } } ``` On-screen text is transcribed into the description, so the slide title is searchable. #### A screen recording with no narration Goal: find a silent bug reproduction by the text of the error dialog. **`request`** ```json title="request" { "query": "\"payment method declined\" dialog", "limit": 3 } ``` The description transcribes UI text. A silent clip has no transcript line. See [Supported File Types](/guides/supported-file-types#video) for the extension table and [Audio](/guides/file-types/audio) for audio-only files.