Importing Transcripts

If you already have a backlog of recorded calls — or you run on a system Omniflow doesn’t host — you can import transcripts in bulk and have them scored against your rubric.

What you can import

Source	How
CSV upload	Drag a CSV with `agent`, `channel`, `started_at`, `transcript`, optional `audio_url`.
Call recording vendor	Connect a vendor (Genesys, Twilio, etc.); transcripts sync nightly.
API	POST conversations to `POST /api/v1/conversations` for real-time grading.
JSONL	Bulk-load NDJSON from your data warehouse.

CSV upload

Open Import

Go to QA → Import → Upload CSV.

Map columns

Omniflow tries to auto-detect column names. Confirm or remap:

Column	Required?	Notes
`external_id`	optional	Lets you cross-reference with your source system.
`agent`	required	Email or external user ID. Unknown agents are auto-created.
`channel`	required	`voice`, `chat`, `email`, `sms`, `social`.
`started_at`	required	ISO 8601.
`transcript`	required	Plain text. Speaker tags (`Agent:`, `Customer:`) are recommended.
`audio_url`	optional	Signed URL Omniflow can fetch for playback.
`tags`	optional	Comma-separated.

Pick a rubric

The rubric to grade against. Use the workspace default or pick a custom rubric.

Run

Imports run in the background. You’ll see progress in QA → Import history and get an email when it finishes.

Vendor sync

For Genesys, Twilio, and other supported vendors, connect once and transcripts sync automatically.

Vendor	Sync frequency
Genesys Cloud	Every 5 minutes
Twilio Voice	Real-time (webhook)
Genesys Engage	Hourly
RingCentral	Hourly
Zoom Phone	Hourly

See Telephony integrations for setup.

API

For real-time grading or bespoke pipelines, POST to /api/v1/conversations:


curl -X POST https://api.omniflow.example/api/v1/conversations \
  -H "Authorization: Bearer $OMNIFLOW_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "external_id": "call_8a2f",
    "agent": "ari@acme.com",
    "channel": "voice",
    "started_at": "2026-05-01T14:23:00Z",
    "transcript": "Agent: Thanks for calling Acme...",
    "rubric_id": "rb_default"
  }'

Response includes the scored criteria within ~5 seconds for short transcripts. See API Reference.

Common pitfalls

Pitfall	Fix
Transcripts have no speaker tags	Add `Agent:` / `Customer:` prefixes — accuracy drops 15-20% without them.
One huge CSV with 100k+ rows	Split into 10k-row chunks — the importer handles them in parallel.
Mixed languages in one rubric	Use language-specific rubrics; same criteria, language-tuned descriptions.

Don’t import raw chat logs unfiltered. System messages, typing indicators, and bot pings inflate token counts and confuse the grader. Pre-process to keep only human turns.

Open in Omniflow

If you want to…	Go to
Connect a call vendor	Telephony integrations
Set up a custom KPI	Custom KPIs
Use the API	API Reference