Embeddings and audio

Vector embeddings, speech-to-text, text-to-speech and translation through the OpenAI-compatible routes.

These surfaces use the OpenAI wire format, so the official SDKs work by changing only the base URL.

Embeddings

curl https://router.triw.io/v1/embeddings \
  -H "Authorization: Bearer $TRIW_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"text-embedding-3-small","input":["first chunk","second chunk"]}'

from openai import OpenAI
c = OpenAI(api_key="sk-...", base_url="https://router.triw.io/v1")
r = c.embeddings.create(model="text-embedding-3-small", input=["first chunk", "second chunk"])
print(len(r.data[0].embedding))

Parameter	Note
input	String or array of strings; batch sizes are model dependent
dimensions	Supported by models that allow truncation (e.g. text-embedding-3-*)
encoding_format	float (default) or base64

Store the model id with the vectors. Embeddings from different models are not comparable, and swapping models later requires re-indexing.

Speech to text

curl https://router.triw.io/v1/audio/transcriptions \
  -H "Authorization: Bearer $TRIW_API_KEY" \
  -F "model=whisper-1" \
  -F "file=@meeting.m4a" \
  -F "response_format=verbose_json"

response_format accepts json, text, srt, vtt, or verbose_json. Use verbose_json when you need segment timestamps. For translation into English from another language, use /v1/audio/translations with the same parameters.

Text to speech

curl https://router.triw.io/v1/audio/speech \
  -H "Authorization: Bearer $TRIW_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"tts-1","voice":"alloy","input":"Your balance is ten dollars."}' \
  --output speech.mp3

The response body is audio bytes, not JSON — write it straight to a file or stream it to the client. voice values and available formats depend on the model.

Choosing models

curl https://router.triw.io/v1/models -H "Authorization: Bearer $TRIW_API_KEY"

Embedding and audio ids appear in the same list as chat models. If an id is missing, that channel is not enabled for your key’s group.

Billing notes

• Embeddings bill per input token; long documents are the main cost driver, so chunk before sending.
• Transcriptions bill per audio minute; TTS bills per input character. Check the model price before batch jobs.
• Failed requests that never reached the upstream are not billed.

Troubleshooting

Symptom	Cause	Fix
400 on embeddings	input empty or too large for the model	Batch into smaller arrays
415 on transcription	Wrong content type	Send multipart form data (-F), not JSON
Audio file is JSON with an error	Request failed; body was the error payload	Check the file’s first bytes
Empty or truncated transcript	File exceeds the model’s duration limit	Split the audio