Every stage as a span
Segmentation, transcription, frame analysis, retrieval and the final summarisation, in order, under one trace for the asset.
By agent type · Video
A video agent segments, transcribes, samples frames, describes, retrieves and then summarizes. By the time the output is wrong it has passed through six models, and the only artifact anyone kept is the last one. Netra keeps all of it as one trace.
Segmentation, transcription, frame analysis, retrieval and the final summarisation, in order, under one trace for the asset.
Video work is mostly a speech and text problem wearing a timeline. The transcript the model reasoned over sits next to what it concluded.
Per-stage timing across a long-running job, so the bottleneck is a number rather than an argument between two teams.
Minutes transcribed, frames sampled and tokens spent attributed to the video and the route that processed it.
Score the summary, the tags or the extracted moments against a reference set — faithfulness to the transcript, correct structure, nothing invented.
OpenTelemetry-native instrumentation means a shot detector or a human QA step joins the trace with a decorator, not a rewrite.
A bad segment boundary makes a perfectly good summariser produce nonsense. Without the intermediate spans, the blame lands on the last model every time.
The moment that mattered fell between two sampled frames. Seeing which frames were actually sent is the difference between fixing the sampler and re-prompting forever.
A pipeline that is fine on a two-minute clip is a budget event on a two-hour recording. Per-asset attribution catches it on the first long file.
Keep the assets that broke as a fixed set and grade every model swap, sampler change or prompt edit against them before it ships. The same evaluators keep scoring production, so drift shows up as a trend rather than a complaint.
Start today
Every stage as a span, cost attributed per asset. Start free.