API Changes – V12.0

Added

Multimodality / input assets:
- POST /v1/input-asset/upload/init – Initialise an input asset upload (presigned S3 URL, input_asset_id).
- POST /v1/input-asset/upload/complete – Complete the upload and start processing.
- DELETE /v1/input-asset/upload/abort – Abort an active upload and clean up the resource.
- PATCH /v1/chat-session/input-assets/ – Bulk update input assets.
- DELETE /v1/chat-session/input-assets/ – Bulk delete input assets.
- PATCH /v1/chat-session/input-asset/{asset_id} – Update a single input asset.
- DELETE /v1/chat-session/input-asset/{asset_id} – Delete a single input asset.
Reasoning control:
- New reasoning_effort parameter in StreamRequest, ChatSession, NewChatSessionSchema and PromptTemplate for client-side control of thinking intensity on supported models (OpenAI o-series, Anthropic Claude 3.7+, Gemini).
Transcriptions (new S3 workflow):
- New three-step upload workflow for transcriptions using S3 presigned URLs (supports files up to 500 MB):
  - POST /v1/transcript/upload/init – Initialises the upload. Returns transcript_id, upload_mode (single or multipart), presigned URL(s), and, where applicable, upload_id and part_urls for multipart uploads.
  - PUT <presigned_url> – Uploads the file (or individual parts) directly to S3 without passing through the API server (URL from the initialisation step).
  - POST /v1/transcript/upload/complete – Completes the upload and starts transcription. For multipart uploads, upload_id and parts (ETags) are required.
  - DELETE /v1/transcript/upload/abort – Aborts an active multipart upload and cleans up the transcript resource. Requires transcript_id and upload_id.

Changed

Chat streaming (/v1/chat/stream/{session_id}):
- Multimodal support: The endpoint now automatically detects unassigned input assets in the session and includes them in the AI request.
- Format change: The response now uses NDJSON (application/x-ndjson) instead of Server-Sent Events (SSE).
- Reasoning: Supports the reasoning_effort parameter for dynamic control of model thinking processes.
Chat providers and reasoning: Replaced static provider variants with a dynamic reasoning parameter (see details).
Input asset management:
- /v1/chat-session/{session_id}/sequence/{sequence_id}/input-asset/{asset_id} now supports PATCH (update) and DELETE (delete), in addition to the simplified /v1/chat-session/input-asset/{asset_id} path.

Deprecated

POST /v1/transcript/ – Direct file uploads for transcriptions are deprecated and will be removed in a future version.

⚠ Breaking change:
The limit is reduced to 50 MB. The new upload workflow is then mandatory for all files larger than 50 MB (see migration).
Early migration to the new workflow is strongly recommended.

Responses from this endpoint include the following HTTP header according to RFC 7234:

Warning: 299 - "This API call is deprecated and will be removed. Refer release notes for details."

Migrating to the new upload workflow:

Replace the previous direct upload with the following three-step workflow:

Step 1 – Initialise the upload:

http

POST /v1/transcript/upload/init
Content-Type: application/json
Authorization: Bearer <token>

{
  "filename": "meeting.mp4",
  "file_size": 52428800,
  "language": "de-DE",
  "diarization": true
}

Response:

json

{
  "transcript_id": "abc123",
  "upload_mode": "single",
  "file_path": "abc123-meeting.mp4",
  "upload_url": "https://s3.amazonaws.com/bucket/meeting.mp4?X-Amz-Signature=...",
  "upload_id": null,
  "part_urls": null,
  "part_size": null,
  "expires_in_seconds": 3600
}

Step 2 – Upload the file directly to S3:

The upload_mode from the initialisation response determines the upload path (single or multipart).

HTTPPython (requests)Node.js (fetch)

http

### Single part (upload_mode: "single")

PUT https://s3.amazonaws.com/bucket/meeting.mp4?X-Amz-Signature=...
Content-Type: video/mp4

<binary file content>

### Multipart (upload_mode: "multipart") – repeat for each part

PUT https://s3.amazonaws.com/bucket/meeting.mp4?partNumber=1&uploadId=...&X-Amz-Signature=...
Content-Type: video/mp4

<binary chunk 1>

# Store the ETag from the response header → pass it to the complete call

python

import requests

# Values from the initialisation response
if upload_mode == "single":
    with open("meeting.mp4", "rb") as f:
        response = requests.put(
            upload_url,
            data=f,
            headers={"Content-Type": "video/mp4"},
        )
        response.raise_for_status()
    parts = None

elif upload_mode == "multipart":
    parts = []
    with open("meeting.mp4", "rb") as f:
        for i, part_url in enumerate(part_urls, start=1):
            chunk = f.read(part_size)
            response = requests.put(
                part_url,
                data=chunk,
                headers={"Content-Type": "video/mp4"},
            )
            response.raise_for_status()
            parts.append({"part_number": i, "etag": response.headers["ETag"]})

# Pass parts to the complete call (see step 3)

javascript

import { createReadStream, statSync } from "fs";

// Values from the initialisation response
let parts = null;

if (uploadMode === "single") {
  const response = await fetch(uploadUrl, {
    method: "PUT",
    body: createReadStream("meeting.mp4"),
    headers: {
      "Content-Type": "video/mp4",
      "Content-Length": String(statSync("meeting.mp4").size),
    },
  });
  if (!response.ok) throw new Error(`S3 upload failed: ${response.status}`);

} else if (uploadMode === "multipart") {
  parts = [];
  const fileSize = statSync("meeting.mp4").size;
  let offset = 0;

  for (let i = 0; i < partUrls.length; i++) {
    const chunkSize = Math.min(partSize, fileSize - offset);
    const response = await fetch(partUrls[i], {
      method: "PUT",
      body: createReadStream("meeting.mp4", { start: offset, end: offset + chunkSize - 1 }),
      headers: {
        "Content-Type": "video/mp4",
        "Content-Length": String(chunkSize),
      },
    });
    if (!response.ok) throw new Error(`Part ${i + 1} failed: ${response.status}`);
    parts.push({ part_number: i + 1, etag: response.headers.get("ETag") });
    offset += chunkSize;
  }
}

// Pass parts to the complete call (see step 3)

No Authorization header is required—the authentication is already included in the presigned URL.

Step 3 – Complete the upload and start processing:

Single part:

http

POST /v1/transcript/upload/complete
Content-Type: application/json
Authorization: Bearer <token>

{
  "transcript_id": "abc123",
  "upload_id": null,
  "parts": null
}

Multipart (ETags from step 2 required):

http

POST /v1/transcript/upload/complete
Content-Type: application/json
Authorization: Bearer <token>

{
  "transcript_id": "abc123",
  "upload_id": "some-upload-id",
  "parts": [
    { "part_number": 1, "etag": "\"abc123\"" },
    { "part_number": 2, "etag": "\"def456\"" },
    { "part_number": 3, "etag": "\"ghi789\"" }
  ]
}

Check status (unchanged):

http

GET /v1/transcript/abc123
Authorization: Bearer <token>

You can retrieve the transcription status with GET /v1/transcript/{transcript_id} (possible values: new, succeeded, failed).

Removed

Instruct API:
- The POST /v1/instruct/http and POST /v1/instruct/stream endpoints have been removed. Please use the corresponding /v1/chat endpoints instead.
Specific reasoning providers:
- Provider variants with fixed reasoning levels in their names (for example, azure-gpt5-high-reasoning-effort or azure-openai-gpt-5_1-low-reasoning-effort) have been removed.
- Use the base provider (for example, azure-openai-gpt-5_4) with the reasoning_effort parameter instead.

Chat providers and reasoning effort

With version 12.0, reasoning intensity (thinking) selection has moved from the provider level to the parameter level. This enables more flexible control without changing the provider.

The `reasoning_effort` parameter

The parameter can be set when creating a chat session or directly in the streaming request (POST /v1/chat/stream/{session_id}).

Permitted values:

Value	UI label	Description
`none`	Off	Disables reasoning or extended-thinking features.
`low`	Standard	Uses a basic reasoning level or moderate thinking budget.
`high`	Verbose	Uses a higher reasoning level or larger thinking budget.

How it works per provider:

The API abstracts provider-specific implementations:

OpenAI (o-series): Maps directly to the native reasoning_effort field.
Anthropic (Claude 3.7+): Controls the thinking budget (for example, 1024 versus 4096 tokens).
Google Gemini: Controls thinking_budget (tokens) or thinking_level.

API Changes – V12.0 ​

Added ​

Changed ​

Deprecated ​

Removed ​

Chat providers and reasoning effort ​

The reasoning_effort parameter ​

API Changes – V12.0

Added

Changed

Deprecated

Removed

Chat providers and reasoning effort

The `reasoning_effort` parameter