Data retention - Chariot API

Chariot keeps two kinds of data for each request: the content you send and receive, and usage metadata about the request. By default, content is kept until it is deleted. A workspace can instead choose a retention period, after which its content expires.

Content and usage metadata

	Content	Usage metadata
Text to speech	Input text and generated audio	Generation id, voice, model, status, character count, credits, timestamps
Speech to text	Audio, transcripts, word timings, display filenames, `custom_id` values	Transcription id, status, mode, language, duration, credits, timestamps
Every request		Which API key or connected app made the request

A retention period applies to content only. Usage metadata is kept so that billing, usage reports, and history listings keep working: an expired generation still appears in history, without its text or audio.

Retention periods

You can choose one of these periods for your workspace: 1 day, 7 days, or 30 days. Content is kept until the period ends, then it expires. Data retention is available on request, generally as part of an enterprise agreement. Contact support@chariot.in to make it available for your workspace. Once it is available, a workspace owner or admin turns it on and chooses the period in Settings > Privacy on the dashboard. Other members can see the setting but not change it. Privacy settings are managed in the dashboard only: API keys cannot read or change them.

What it applies to

The retention period applies to requests made with an API key and to requests made by apps you connect to your workspace through the Chariot MCP server. Work you do while signed in to Chariot, such as in Studio or the Playground, is kept until you delete it, whatever the retention setting.

New requests only

The period is fixed when Chariot accepts a request. Changing the setting affects new requests only: it never shortens or extends how long existing content is kept. If you turn retention off, new requests are kept until deleted, and earlier content still expires on its original date.

When content expires

Resources under a retention period carry a content_expires_at timestamp. It is the time the request was accepted plus the period. A null value means the content is kept until deleted.

Resource	Field
TTS history (`GET /v1/tts/history`)	`content_expires_at` on each entry
TTS audio links (`GET /v1/tts/history/{inference_id}/audio`)	`content_expires_at`
Transcriptions (`GET /v1/asr/history`, `GET /v1/asr/transcriptions/{transcription_id}`)	`content_expires_at`

As soon as content_expires_at passes, the API stops returning the content:

Text to speech: tts_input_text is null, has_audio is false, and the audio endpoint returns 410.
Speech to text: filename, custom_id, transcript_preview, result.text, and result.words are null. Status, duration, and billing remain.

Chariot then deletes the stored content in the background. A recovery copy of stored files is kept for about 7 more days, then deleted permanently.

Retention periods of 1 to 30 days

Content is available through the API until content_expires_at.

Text to speech: history entries keep their text, and you can fetch the audio by generation id until it expires.
Speech to text: transcripts and word timings stay available from the transcription detail. For a synchronous file sent with input_audio_url, Chariot keeps only a display filename, never the URL you sent, because signed URLs carry credentials.
Asynchronous batches: results expire a period after the batch is accepted, not after it finishes. Fetch results before content_expires_at.

If Chariot cannot read your workspace’s retention setting when you submit a batch, POST /v1/asr/batches returns 503. Retry shortly.

Fetch audio by generation id

Every TTS generation has an id: the Inference-Id header of POST /v1/tts, the X-inference-id header of POST /v1/tts/stream, or inference_id in TTS history. Keep that id, and request a short-lived link to the audio whenever you need it:

# JSON with a signed link
curl https://api.chariot.in/v1/tts/history/3f2c9a71-5b0e-4d7a-9c1e-8a4b6d2f0e13/audio \
  -H "chariotai-api-key: $CHARIOT_API_KEY"

# Or follow the redirect and save the WAV file
curl -L "https://api.chariot.in/v1/tts/history/3f2c9a71-5b0e-4d7a-9c1e-8a4b6d2f0e13/audio?redirect=true" \
  -H "chariotai-api-key: $CHARIOT_API_KEY" \
  --output speech.wav

import os
import requests

API = "https://api.chariot.in"
headers = {"chariotai-api-key": os.environ["CHARIOT_API_KEY"]}
inference_id = "3f2c9a71-5b0e-4d7a-9c1e-8a4b6d2f0e13"

link = requests.get(f"{API}/v1/tts/history/{inference_id}/audio", headers=headers)
link.raise_for_status()
audio = requests.get(link.json()["url"])  # no API key: the link is signed
audio.raise_for_status()
with open("speech.wav", "wb") as f:
    f.write(audio.content)

{
  "inference_id": "3f2c9a71-5b0e-4d7a-9c1e-8a4b6d2f0e13",
  "url": "https://storage.example.com/tts_audio.wav?signature=REDACTED",
  "url_expires_at": "2026-11-02T10:15:00Z",
  "content_expires_at": null
}

The link works for about 15 minutes, and never past content_expires_at. Request a new one instead of storing it.
Anyone with the link can download the audio until it expires. Do not log, cache, or share it.
Audio is stored after the response finishes. A request made right after a generation can return 404 for a few seconds.
404 means the generation is not in your workspace or has no stored audio. 410 means its content has expired.
You can request up to 600 links per workspace per minute. Over that, the endpoint returns 429.

Permanent audio URLs are being retired. The Audio-Url and X-audio-url response headers and the audio_file field in TTS history can contain permanent links today. They are being retired in favor of the audio endpoint, and the retirement date will be announced by email. Store the generation id, not the URL.

WebSocket audio and `enable_logging`

Your workspace’s retention setting decides what is kept, on every endpoint. On the two TTS WebSockets (/v1/tts/ws and /v1/tts/ws-multi), enable_logging=false keeps a session’s audio out of storage; each segment’s text is still recorded in your TTS history and follows your retention setting. If you omit it, or send true, the retention setting decides for the audio too.

Model training

Content from your workspace may be used to improve Chariot’s models. To turn it off, a workspace owner or admin switches off Help improve Chariot’s models in Settings > Privacy. If your workspace does not show the setting yet, contact support@chariot.in and Chariot will make it available. Content can be used only if the setting was on when the content was created and is still on. Turning it off excludes your content from then on, including content created earlier.

What retention does not cover

Usage metadata, such as ids, status, duration, credits, and timestamps. It is kept for billing and usage reporting.
Webhook payloads. Transcription webhooks carry ids and a result path, not transcript text. See Webhooks.
Copies you make. Audio and transcripts you download or save are yours to manage.

​Content and usage metadata

​Retention periods

​What it applies to

​New requests only

​When content expires

​Retention periods of 1 to 30 days

​Fetch audio by generation id

​WebSocket audio and enable_logging

​Model training

​What retention does not cover

Content and usage metadata

Retention periods

What it applies to

New requests only

When content expires

Retention periods of 1 to 30 days

Fetch audio by generation id

WebSocket audio and `enable_logging`

Model training

What retention does not cover