video-digest-segment

video-digest-segment

LLM stage — no scripts. Input: <item_dir>/transcript/timeline_30s.txt (from ingest). Output: <item_dir>/talks.json.

This is the fragile stage — read this before running

A single pass over a braided feed (stream cutting between stages, re-shows, MC filler) drops real talks. Always run the verification pass below.

Procedure

1. If an authoritative schedule exists (conference program page, description chapters, chapters.txt), save it to <item_dir>/schedule.txt — titles/speakers/times in order. It's reconciliation ground truth. No schedule → derive talks purely from content.
2. Run the segmentation prompt: references/segment-prompt.md (fill {ITEM_DIR}, {CONTEXT}, paste schedule facts). One agent for the whole stream.
3. VERIFY (mandatory): walk the schedule (or the talk list you produced) in order; for EACH expected talk, grep transcript/transcript.tsv for the speaker's self-intro or a distinctive title phrase and confirm a matching entry exists in talks.json. A scheduled talk braided with cutaways must still appear (kind main/talk) even if windows overlap. Timestamps drift — trust transcript CONTENT over scheduled clock times.
4. Sanity-check output: talks ordered by start; start/end "HH:MM:SS"; leading applause trimmed; kinds valid.

talks.json schema

{"talks":[{"id":1,"start":"00:12:30","end":"00:31:05","slug":"kebab-slug",
  "title":"...","speaker":"Name · Org","kind":"main",
  "official_time":"9:15am|","confidence":"high|med|low","note":"self-intro quote / ambiguity"}]}

kinds: main/keynote (headline), fireside, track/talk (regular), cutaway (feed cut to other content — excluded from digest), nonspeech (excluded). Exclude pure breaks/pre-roll entirely; dedupe re-shows.