Skip to main content

Summarize a clip using Tagstore

GET 

/q/:hash/rep/summarize_v2

Generates a structured summary for a time-bounded clip of a content object by fetching tags from the Tagstore service. Unlike /rep/summarize, this endpoint supports multiple quality levels that select different LLM backends and prompt strategies, enabling a trade-off between latency and output richness.

Quality levels:

LevelLLM backendUse case
testLightweight testing modelFast iteration during development
defaultLightweight Ollama modelProduction general-purpose summaries
highQwen3 thinking model (vLLM)High-quality, reasoning-heavy summaries

The prompt is automatically adapted based on the content title metadata and detected persons (cast references) associated with the hash.

A configurable padding (in milliseconds) is applied symmetrically around start_time and end_time before querying Tagstore, so context at clip boundaries is not lost.

Request

Responses

Clip summary generated successfully.