video-summarization

To summarize a video and extract specific events, you provide the clip and define a scenario with required events and optional objects of interest. The service routes short clips to VLM and long clips to LVS to generate a narrative summary and structured event timeline.

For summarizing long videos over 60 seconds, the service routes requests to the Long Video Summarization microservice `/summarize` endpoint. It probes LVS readiness and explicitly falls back to VLM-only summarization if the LVS API is unreachable.

Yes, you can use VLM for video summarization on clips under 60 seconds. The service automatically routes short-video requests to the VLM OpenAI-compatible chat completions endpoint to generate clear narrative summaries.

Video summarization requires mandatory human-in-the-loop confirmation for prompt and scenario selection before processing. This ensures the extracted events, objects of interest, and scenario parameters match your specific operational requirements.

The system handles LVS microservice readiness by probing the service before routing long video summarization requests. If the probe determines LVS is unreachable, it cleanly falls back to VLM-only summarization to ensure the task completes.

When using VLM fallback for long video summarization instead of LVS, the primary limitation is a potential reduction in long-form reporting quality. LVS provides higher-quality structured event extraction for clips 60 seconds or longer compared to VLM.