Feature guide

Vocabulary and Patterns

Give Omi the exact encounter terms and structured token shapes that matter for this recording. Both features are constrained by safety rails and are never allowed to become a free-form prompt.

Custom vocabulary

Pass an exact list through vocabulary. Good entries are patient names, medicines, devices, procedures, abbreviations, clinic names, and other terms likely to be spoken in this encounter.

curl https://api.omi.health/v1/audio/transcriptions \
  -H "Authorization: Bearer $OMI_API_KEY" \
  -F [email protected] \
  -F model=omi-medical-1 \
  -F language=en \
  -F response_format=verbose_json \
  --form-string 'vocabulary=["Tinel","Hepcludex","tirzepatide"]'

Use an encounter list, not a dictionary dump

Focused lists of 5–30 terms outperform large generic dictionaries. Include only terms you expect in this visit. Do not send prose, instructions, or a transcript guess.

How vocabulary is applied

Effective list sizeModeBehavior
1–30 termsdirectUses the validated direct context path with structural safety checks.
31–1,000 termsretrievalSelects acoustically relevant candidates first, then applies context.
Arabic or Hindidirect onlyMaximum 30 terms because the fast multilingual draft does not cover these languages.
language=autoper detected languageMaximum 30 effective terms per language; stored language view is selected after detection.

Request terms are merged above the API key’s stored, language-specific dictionary. Set dictionary=false to suppress stored terms for one request without deleting them.

Safety and audit

Every vocabulary request runs a suspect-insertion rail. Direct mode checks number changes, runaway length, copied prompt text, and repeated context terms. Retrieval mode keeps the acoustic-support rail. A suspicious result falls back to the no-context transcript.

{
  "vocabulary": {
    "enabled": true,
    "mode": "direct",
    "requested_terms": 3,
    "applied_terms": 3,
    "suspect_insertions": 0,
    "safety_fallback": false,
    "language": "en"
  }
}

Use the audit to monitor applied terms and fallbacks without logging transcript content. Omi records tenant-scoped counts for mode, requested/applied terms, suspects, and safety fallback.

Patterns v0-preview

Patterns are per-request hints for structured tokens such as dosage forms, order codes, and identifiers. Only set patterns you expect in the current audio. Omi applies formatting only when the writer independently supports the same token.

curl https://api.omi.health/v1/audio/transcriptions \
  -H "Authorization: Bearer $OMI_API_KEY" \
  -F [email protected] \
  -F model=omi-medical-1 \
  -F language=en \
  -F response_format=verbose_json \
  --form-string 'patterns=[0-9]{1,3}(mg|ml|mcg),[A-Z]{2}-[0-9]{4}'
  • Up to 32 expressions; each is validated before inference.
  • Supported building blocks include [0-9], [A-Z], bounded repeats, literal groups/alternatives, letters, digits, spaces, and hyphens.
  • General-purpose regex operators and unbounded expansion are rejected with 422.
  • Patterns are never saved to the tenant dictionary or applied globally.
  • Patterns are currently enabled for invited preview keys.

Vocabulary and Patterns affect the final writer only

In realtime sessions, vocabulary can improve corrected display text and the official record. It does not bias the disposable live draft. Patterns remain a final-output feature and never rewrite transcript text without independent writer evidence.