Bulk reprocess calls stored without a transcript #7

Open
opened 2026-08-19 22:06:54 -04:00 by logan · 1 comment
Owner

Calls recorded before the gs:// fallback fix (a2cd2c5) were stored while STT was silently disabled, so they have no transcript. No transcript means no extraction, no correlation and no incident — those calls are permanently invisible to the product despite the audio existing.

Needs a bounded, resumable bulk reprocess: select calls that have audio but no transcript, re-run the pipeline, rate-limited so it cannot exhaust the OpenAI balance in a single pass.

Mind the cost exposure — see the OpenAI billing issue. This should only run with a spend ceiling decided in advance.

Calls recorded before the `gs://` fallback fix (`a2cd2c5`) were stored while STT was silently disabled, so they have no transcript. No transcript means no extraction, no correlation and no incident — those calls are permanently invisible to the product despite the audio existing. Needs a bounded, resumable bulk reprocess: select calls that have audio but no transcript, re-run the pipeline, rate-limited so it cannot exhaust the OpenAI balance in a single pass. Mind the cost exposure — see the OpenAI billing issue. This should only run with a spend ceiling decided in advance.
Author
Owner

Scope grew. The AI-off window of 2026-08-19 23:05 to 2026-08-20 02:14 UTC added roughly 2,322 more calls with audio but no transcript (see CORRELATION_REVIEW.md). These are invisible to the product for the same reason as the pre-a2cd2c5 calls.

This also means any correlation analysis over a window that includes an AI-off period will read those calls as orphans when they were never processed at all. The review agent hit exactly that: 246 of 250 sampled orphans had no transcript, making the orphan miss-rate undeterminable for that window. Worth having the debug view distinguish never-processed from failed-to-link.

Scope grew. The AI-off window of 2026-08-19 23:05 to 2026-08-20 02:14 UTC added roughly 2,322 more calls with audio but no transcript (see CORRELATION_REVIEW.md). These are invisible to the product for the same reason as the pre-a2cd2c5 calls. This also means any correlation analysis over a window that includes an AI-off period will read those calls as orphans when they were never processed at all. The review agent hit exactly that: 246 of 250 sampled orphans had no transcript, making the orphan miss-rate undeterminable for that window. Worth having the debug view distinguish never-processed from failed-to-link.
Sign in to join this conversation.
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: logan/server-26#7