Frames only (requires a frames spec)

Present only when the caller passed the translate param. An LLM translation of the transcript into the requested ISO-639-1 language, billed per audio-minute like transcription.

LLM summary of the transcript (present only when the digest field was requested and a transcript was available).

A spoken-audio rendering of the video's digest (present only when "audio_digest" was requested and a digest was available to synthesize). Never served from the content-hash cache — its signed URL expires and the voice can vary per request. Billed per estimated spoken second (from the digest text length, ~15 chars/s) ONLY when the mp3 is actually produced; a TTS/upload failure omits this field, adds a warning to the top-level warnings array, and is not charged — the rest of the extraction stays usable. Failure reasons include no TTS provider configured (analogous to the worker-internal TTS_UNAVAILABLE code: neither OPENAI_API_KEY nor GROQ_API_KEY set) and an unsupported language on the only configured provider (TTS_LANG_UNSUPPORTED: Groq/PlayAI only speaks English/Arabic — set an OpenAI key for other languages); neither surfaces as an HTTP error code, only as a warnings[] string.

Typed, structured understanding of a video (present only when "structured" was requested): the transcript plus up to 12 sampled keyframes are run through a vision LLM (gpt-4o-mini, temperature 0) that returns chapters, entities, products shown, claims, and key moments as strict JSON. Deterministic and cacheable — unlike audio_digest it carries no expiring signed URL, so it takes part in the normal content-hash cache. Best-effort like digest: any download/VLM/parse failure omits the whole field, adds a warning to the top-level warnings array, and is NOT charged; the rest of the extraction stays usable. Worker-internal failure codes (never surfaced as an HTTP error, only as a warnings[] string): STRUCTURED_UNAVAILABLE (no vision provider configured — OPENAI_API_KEY unset), STRUCTURED_PROVIDER_ERROR (the provider returned a non-2xx response), STRUCTURED_PARSE_ERROR (the model's output could not be parsed/validated into this schema). Billed a flat $0.03 per video (perStructuredVideoMicros, floor-pinned — see /pricing) charged only when the analysis is actually produced.

Up to cap_applied TOP-LEVEL comments for a video, sorted by the platform's "top" ranking (present only when "comments" was requested). Supported on YouTube only (via yt-dlp's info-json comments). Reddit's own public comment source — the unauthenticated thread JSON API — worked the same way until Reddit deprecated unauthenticated access to it on 2026-05-28; TikTok/Instagram/Pinterest never had a reliable public comment source. Billed a flat $0.0045 per call (independent of comments_cap), charged only when comments were actually produced. Best-effort: any fetch/parse failure, or an unsupported platform, omits the whole field with a warning and is NOT charged — never surfaces as an HTTP error code. Worker-internal failure codes behind those warnings (never an HTTP error, only a warnings[] string): COMMENTS_UNSUPPORTED (the platform has no comment source — drop "comments" from fields for it; see GET /v1/platforms), COMMENTS_PROVIDER_ERROR (the yt-dlp info-json fetch or parse failed).

Aggregated LLM sentiment over a video's comments (present only when "comment_sentiment" was requested, which auto-includes "comments" as its input): ONE Groq chat call (llama-3.3-70b-versatile, temperature 0, JSON mode) reads up to 100 comments and returns the audience mood AS A WHOLE — not per-comment scoring. Requires at least 5 fetched comments; fewer (or a model/parse failure) omits the whole field with a warning and is NOT charged — never surfaces as an HTTP error code. Billed a flat $0.006 per call, on top of the $0.0045 "comments" call it rides on. No competitor offers this rollup.

Change since the most-recent prior fetch of the same url (present only when the delta field was requested). First time a url is seen -> { hasPrior: false } and nothing else. A per-hour rate is null when a counter was unavailable or no time had elapsed.

On-screen text per frame (OCR), same index as frames.