Insights

The output of running an intent over a recording — searchable, replayable, and traceable.

An insight is a single execution of an intent against an audio. It links the source recording, the analysis output, and any downstream integration delivery into one record.

Insight Lifecycle

Each insight moves through these statuses:

Status changes are streamed to the dashboard in real time, so a long-running insight updates without a refresh.

Insight Detail Page

/insights/[id] shows:

• Source audio — link to the recording, duration, channel context
• Intent that was run, including the prompt
• Output — the rendered LLM result (Markdown supported)
• Integration logs — for each delivery: provider, status, external URL, execution time
• Replay — re-run the same intent against the same audio (useful after editing the prompt)
• Confidence — when the model returns a self-rated confidence score

Filtering and Sorting

The insights index supports:

• Status filter — all, pending, processing, completed, failed
• Sort — newest, oldest, by confidence, or surface failed runs first
• Search — across audio file name, channel name, and intent name
• Live counters for processing and failed runs at the top of the page

Re-running

Re-running an insight creates a new insight record rather than mutating the original. This preserves history when you tweak prompts and want to compare outputs.

Failure Handling

When an insight fails:

1. The status flips to failed and the error surfaces on the detail page
2. The integration log records the provider's error message (if delivery was the failure point)
3. You can edit the intent and re-run; nothing is auto-retried beyond the Trigger.dev task's retry budget

Best Practices

• Sort by failed to triage broken intents quickly
• Use the replay button after prompt iteration instead of re-uploading audio
• Set up a download intent alongside critical integrations as a fallback artifact