What a recap actually is
A recap is a compressed retelling of a long-form video. The compression ratio matters — a 4-hour cricket match compressed to 90 seconds is a different artifact than the same match compressed to 10 minutes. A recap layer must support the full range and let the operator choose.
Inside the artifact, three things must be true: every claim must be sourced (a citation-backed answer pattern), the salience selection must reflect domain semantics (a wicket is more salient than a forward defensive stroke), and the language and tone must match the audience (Hindi commentary recap for IN, English for international).
How salience is determined
Deepgrip combines three signals to identify high-salience moments in a recording:
- Entity events: named entities that appear in domain-specific event roles (a wicket, a goal, a milestone run, a named guest entering, a motion being introduced in parliament).
- Audio cues: applause, crowd noise spikes, music beds, dramatic silences. These are surprisingly reliable salience markers.
- Linguistic markers: rhetorical patterns, emphatic phrasing, repetition, named-entity density.
Output shapes
A recap can be rendered as: a 90-second video with auto-cut clips and captions; a 3-minute structured summary with embedded timestamped citations; a 10-minute recap with longer clips and full surrounding context; or a JSON timeline that downstream editorial tools (NLEs, CMS plug-ins) can consume.
The same source video produces all of these from one indexing pass. Recaps in 121 languages from one source are shipped without re-editing — the translation pipeline aligns at the segment level.
When to use recap vs compile
A recap summarises a single recording. A compile assembles moments across many recordings along a chosen dimension (one player across a tournament, one topic across a season). Both are valuable, both are outputs of the same searchable archive — but they answer different editorial questions.
For one match, one episode, one session — use a recap. For "every time speaker X discussed topic Y across the year" — use a compile.
Editorial guard rails
A recap is shipped to an audience. Editorial guard rails matter: muted-term lists, brand-safe overrides, attribution requirements, approval queues. Deepgrip's recap pipeline supports all of these as first-class settings, configurable per archive and per output channel.