FalconVQA Docs
Using FalconVQA

The video detail page

Everything stored about one video — overview, chunks, speech, entities, and the raw record.

/projects/{projectId}/videos/{videoId} is the explorer for a single video. Where the chat answers questions, this page shows what the answers are made of.

It loads from GET /api/videos/{id}/details, which returns the row, core's own metadata, every chunk with every analyzer's output attached, and every stored aggregate. Chunks are paged out of core 500 at a time and walked until core's total is satisfied — core paginates because an agent reading chunks wants a few; this page is the opposite case.

Sources settle independently. A video analysed without people has no entities aggregate; a video mid-ingest has no chunks yet. Neither is a failure, and neither blanks the parts that did load — per-source errors are listed and everything else still renders.

Tabs

Overview

The video-level picture: player, poster, duration, chunk count, the analyzers that ran, and the aggregates that exist. This is where summary, chapters, events, stats and novelty are rendered — tiered summaries with the finest tier first, chapters as titled spans, events as a timestamped list, stats as counts and extremes.

Chunks

Every chunk in order, with each analyzer's output for it: the scene description and its facets, on-screen text, detected objects, and the people seen. This is the raw material of every search hit, shown unabridged.

Speech

The transcript, attributed when the video was analysed with diarization. Alongside it, speaker_stats (talk time, turns, handovers, share per speaker) and sentiment when those aggregates ran.

Entities

People linked across the whole video: each one's description, the narrative of what they did, first and last seen, appearance count, and dwell time from entity_timelines. cooccurrence shows which people appeared together.

Only present when the video was analysed with people.

Raw

The stored record as returned, for when you need to see exactly what an analyzer produced rather than a rendering of it.

Re-running aggregates

The page can re-run video-level passes without re-analysing anything, through POST /api/videos/{id}/aggregates. Aggregators read stored analyzer output rather than the video, so this is the cheap half of re-processing.

Five aggregators bill LLM calls — summary, chapters, events, entities and object_entities. That is why re-running is a deliberate action rather than something the page does on load.

Deleting and re-indexing

Both are available here as well as from the project grid, and behave identically — see Adding videos.

On this page