AI recaps

When a meeting ends, Olivasal transcribes it and writes a summary you can actually read. Recaps run on our managed AI by default, or on your own GPU so transcripts never leave your network.

What it produces

• Transcript — speaker-attributed (via diarization), with per-segment timing.
• Summary — rich headings plus a visual flowchart of how the meeting unfolded.
• Action items — extracted as a checklist.
• Searchable library — every recap is indexed by summary, transcript, action items, and room name.

How it works

1. The recap is sourced from the recording — so AI recap requires recording to be on. (It is therefore unavailable under E2EE.)
2. Transcription runs on our speech-to-text engine. Language is auto-detected per track; Tamil routes to a Tamil-optimized speech model, everything else to the default engine. Original language is preserved — Olivasal never translates the transcript.
3. Summarization runs through a local language model (configurable via the SUMMARY_MODEL env var).
4. Status surfaces live: transcribing → summarizing → ready. Failures show a reason — no silent breakage.

A serial queue runs one transcription at a time (concurrent passes OOM the GPU), and recaps orphaned by a restart are re-enqueued automatically.

Regenerate

Regeneration is append-only: a new version is added to the history while the previous one stays visible until the new one is ready. The latest version is mirrored to the top of the meeting detail page.

Status visibility

Polling reflects the two-phase commit — the transcript appears first, then the summary and action items. Add a language by setting it once in the language map; no other code changes are needed.