Pegasus 1.5 brings timestamped video metadata to Bedrock
AWS’s October 19 guide explains how teams can define video segments and receive structured results from TwelveLabs’ model, whose Bedrock availability was announced on October 9.
The AWS walkthrough illustrates the task with a broadcast: define its editorial sections, then request titles, summaries and topics for them. Results pair the requested information with beginning and ending timestamps. AWS describes a workflow in which people review and refine the generated descriptions, retaining a human editorial role.
Access has a consequential boundary. The current Bedrock model card specifies cross-Region inference: requests must use either the US geographic profile or the global profile. It explicitly excludes in-Region on-demand inference, so this deployment does not offer the option of keeping inference within one selected AWS Region.
The response also reports processing conditions. Its definition_errors field lists segment definitions that could not be processed; an empty list means all definitions were applied. Separately, stopReason set to length indicates that generation reached the output-token limit.
Language coverage also matters for choosing footage. TwelveLabs’ specification lists full English support and partial support for twelve other languages, including French, German and Spanish. Hungarian is absent from that published list.
The developer’s context-window description gives another planning constraint: the request and response share one capacity budget. Video, audio transcription, instructions, reference images, an output schema and segment definitions all count alongside the generated answer. The available capacity therefore covers more than the footage alone.
Sources(3)
- AWS walkthroughChecked:
- Bedrock cardChecked:
- Model specificationChecked: