Skip to main content
Recap turns a completed transcript into a summary, action items, next steps, and structured fields. It accepts transcripts from Hear, Omni, or your own system. Recap does not transcribe audio; use Hear first when your input is a recording.

Settings and customization

Speaker diarization before call analysis

If you start with a recording, use Hear speaker diarization to obtain a speaker-labelled transcript before preparing Recap input. Speaker labels are neutral: confirm agent/customer roles from your call context before mapping them into Recap. If you already have a suitable transcript, you can submit it directly.

Enable Recap

Use recap:lookup to list or read existing summaries in the organization and recap:create to submit a transcript. Changing the organization setting requires recap:configure. Existing keys with the legacy recap:read permission can still look up and create summaries.

Create a Recap from an async transcription job

Add a stable call_id to the job. The other Recap fields are optional. language controls the language of the notes, not speech recognition. For speech recognition, use the async job’s transcription_language, or omit it for automatic detection. See STT language and writing system.
When transcription completes, PyAI sends the complete segment list to Recap. The Recap record moves through pending, processing, and complete, or failed if processing cannot finish. pending is the instant before the record reaches the Recap service, so what you actually read back is processing and then complete: POST /v1/recap/calls/{call_id} answers 202 with "status": "processing", never "pending". Poll GET /v1/recap/calls/{call_id} until it is complete or failed.

Create a Recap from Hear streaming

For a Recap-enabled organization, the streaming adapter submits committed utterances when the socket closes. Pass a stable call_id so the result can be joined to your own call record:
Keep sending audio frames through pauses so turn detection can observe silence. Only committed utterances become part of the retained transcript.

Submit an existing transcript

If you already have speaker-labelled utterances, trigger Recap directly:

Read the call detail

A completed response includes two separate artifacts:
  • transcript: the full retained utterance list with speaker roles and timing.
  • record: typed intelligence (recap.record.v1). tldr is the one-line headline; summary is the detailed notes. action_items, next_steps, talk_ratio, signals, and fields are always present (null or empty when unknown).

Choose a pack

default_pack_id on PUT /v1/recap/config is the organization fallback. pack_id on a Hear job, Hear stream, or manual Recap submission overrides it for that call. A pack selects the field and signal definitions Recap should produce for a use case. For example, sales_outbound targets outbound sales notes. It does not change speech recognition or speaker separation. Keep a stable pack id in your application rather than inferring one from call direction.

Receive completion webhooks

Set webhook_url on the organization Recap config:
When processing completes, PyAI posts a recap.complete payload containing call_id, pack_id, status, completed_at, and the typed record. Use the shared signature verification guide. Treat delivery as retriable and deduplicate using the call identity and event semantics.

Salesforce field mapping

PUT /v1/recap/crm-config configures the design-partner Salesforce mapping. It maps logical Recap fields to Salesforce API field names and can create a Task from the TL;DR and summary. Secret fields are redacted on reads; omit an existing secret on update to preserve it. Most teams should configure Salesforce from Console → Integrations and use the Salesforce integration guide. Use the direct CRM-config endpoint only when you explicitly need the API-managed field map.

Review in the console

The Recap call view keeps processing and failed calls visible. A completed call has separate Summary, Insights, Transcript, and Ops views. The Transcript view contains every retained utterance and can be copied with timestamps. Recap rows, including their transcript and generated intelligence, are retained for 90 days. Input audio follows the retention policy of the Hear surface that created the transcript. See Security and data handling.

Troubleshooting

Topics, moments and follow-ups

The Recap overview has an action worklist and a Topics view. Filter actions by status, owner or text, then export the matching rows. Topic counts compare the selected 7, 30 or 90 days with the preceding period of equal length. Matching names are grouped; each call counts once per topic. Reports include all matching calls, not just the current table page. Open a call workspace to review Moments & follow-ups beside its transcript. Topics, action items, decisions and risks can link to a supporting passage. Extraction uses verbatim quotes and transcript turn references. Unsupported or ambiguous references stay unlinked instead of using an estimated timestamp. Click a linked moment to reveal the passage and play its recording when present. Use Tag on a transcript turn to add a topic, action, decision, risk or note. Assign actions to a teammate, set a due date, add tags and mark them completed. Changes are shared with the organization and saved with revision history. Owners, admins and developers can edit; billing members have read access. Completed by team records a teammate’s update, not confirmation from a CRM or scheduling system. The transcript and original Trace scorecard stay intact. Follow-up reports use the call’s received date. Tasks from calls outside the selected period are not included. Literal due phrases from extraction remain visible; overdue counts use a date selected by a teammate.