YouTube extract

Open /youtube-extract, paste a public YouTube URL, and select Create extract · Live Jev. The app retrieves an existing caption track and asks Jev to score each natural chunk. It never calls a text-generation, transcription, or translation model.

Retrieval and limits

The server reads YouTube watch metadata, then uses its public player and timedtext endpoints. These are unofficial interfaces and may change or block server IPs. The official YouTube captions download API requires authorization; this prototype does not claim universal video coverage. Public player discovery follows the approach documented by youtube-transcript-api.

English captions are preferred, with manual tracks before automatic ones. If English is unavailable, an existing track in its source language is used, again preferring manual captions. Their origin is displayed. No captions means an error, never invented text. Requests use fixed YouTube hosts, disallow redirects and translated tracks, limit upstream bodies to 2 MiB, and share a 20-second abort deadline. No cookies or credentials are forwarded to YouTube. Transcript limits are 5,000 caption lines, 60,000 characters, 200 natural chunks, and 6,000 characters per natural unit. Oversized tracks fail without truncation.

Decisions and source fidelity

Caption gaps over 1.5 seconds, terminal punctuation, and speaker markers create natural boundaries. A chunk retains original text, caption line indices, and start/end times. Multiple sentences in one caption share that caption's timestamp; these are source cue times, not inferred word timings. Unpunctuated runs remain whole until a natural boundary. Grammatical completeness is best effort.

Cleanup removes hesitation tokens, comma-delimited filler phrases, and immediate hyphenated false starts. Meaningful uses such as “I like ice” or “you know the answer” remain. Cleanup uses English filler heuristics, which can misread meaningful words, especially in other languages. Inspect original captions alongside cleaned text.

Each nonempty chunk uses one /api/run request containing a score relevance question and a yes/no choice key-claim question. The complete bounded transcript and video title provide topic context; source text is marked untrusted. Relevance and confidence must be finite values in [0,1]. Missing, invalid, failed, or cancelled results remain unknown; incomplete scoring stops the run. No output text is taken from a model response.

Selection excludes incomplete scores and relevance below 0.5, then sorts by descending relevance and stable source order. Lower-ranked chunks with word-set Jaccard overlap greater than 0.6 with a retained higher-ranked chunk are duplicates. It keeps up to $\text{ceil}(\text{total} \text{chunks} \times \text{length} \text{percentage} / 100)$ and restores chronological order. Key-claim flags are metadata, not a filter. Raising length cannot restore duplicates or below-threshold chunks.

The brief permits an empty connector and also requires every output word to trace to the transcript. This version uses the empty connector rule: existing source transitions stay, and whitespace joins the cleaned passages. It does not invent “Next” or “Finally,” or rewrite text to improve flow. The result is labeled an extract, not a true summary.

When it fails

A failure names one cause rather than listing what might have gone wrong. The response carries cause, a summary, the detail behind it, a fix, and — when a bound was crossed — the limit that was hit with the value measured.

Cause	Means
url_invalid	The address did not parse as a YouTube video link; no request was made.
video_unavailable	404 or 410: private, deleted, region-blocked, or a wrong id. Captions were never reached.
rate_limited	429 against the server's address, not your key or account.
youtube_error	Another non-success status from a fixed YouTube host.
page_too_large	The watch page passed the 2 MiB read bound.
metadata_missing	The page exposed no player key and no initial player response.
metadata_unreadable	Player metadata began but did not close as balanced JSON.
no_caption_tracks	The video genuinely publishes no captions. Nothing is transcribed as a fallback.
track_rejected	A listed track's URL was not a plain timedtext address on www.youtube.com.
captions_empty	The track fetched but held no readable cues.
timed_out	The shared 20 second retrieval budget elapsed; no partial track is used.
cancelled	Stopped before captions were read.
redirected	A redirect was refused rather than followed.
limit_lines	Over 5,000 caption lines.
limit_characters	Over 60,000 characters.
limit_chunks	Over 200 natural chunks; each chunk is one scored request.
limit_unit	One chunk over 6,000 characters, which an unpunctuated run can reach alone.

limit_chunks and limit_unit were previously one message that named both, so a reader could not tell which bound was crossed. They are separate causes and report the measured value.

The limits are listed in the workspace under Limits before a run, from the same source the checks use, so the stated numbers cannot drift from the enforced ones.

Inspecting a run

The length slider recomputes locally without Jev requests. The raw list shows kept/dropped reasons, scores, key claims, confidence, cleaned text, originals, and source links. Kept passages link directly to their video timestamps. Human comparison with both kept and dropped source passages is the verification method; neither confidence nor extraction proves completeness.

Calls attempted, processed/total chunks, kept count, mean kept relevance, elapsed time, and running input-cost estimates update during scoring. Pricing uses the app's shared estimate function and reported input tokens; missing usage and in-flight/failed calls remain unknown. Input-only cost excludes output and hosting and is not an invoice. The shared usage dashboard retains its existing key and rate-limit behavior.

Stop, unmount, or a personal-key change cancels the run and prevents queued calls. Partial results remain labeled partial. Only URL and length persist as a browser draft; results and keys are not stored in that draft. Changing the URL clears the old result. Live scoring sends the bounded transcript to TypeSafe; shared-key deployment still needs the controls in deployment.md.

Automated unit/API and browser tests use synthetic captions and mocked Jev responses; no shared credits are consumed.