Version 2.4.5#

New release of Sherpa AI Server: Version 2.4.5.

1. Expanded speech recognition configuration#

Administrators can connect external Whisper-compatible services using Bearer token authentication or a token in the request parameters, and can also configure the service response format. If needed, audio recordings in WebM and OGG formats are automatically converted to WAV before being sent.

This makes it easier to integrate Sherpa AI Server with external speech recognition services.

Important for external monitoring: the GET /api/v1/audio/health method has been removed. Availability checks that used this route must be reconfigured.

2. Expanded branding options#

The portal name on the sign-in screen, in the side menu, and in the browser tab title is now taken from the branding settings. For customer deployments, localized overrides of interface texts are also supported.

The administrator can completely hide the language selector on the sign-in screen and in the header by using the hide_language_selector parameter.

This helps adapt Sherpa AI Server more precisely to the organization's brand identity and language requirements.

3. Updated placement of pinned sections#

Pinned sections now appear at the end of the main menu next to chats and assistants. The permanent “Assistants” section cannot be pinned again.

This makes the side menu more compact and avoids duplicate sections.

4. Fixed bulk deletion#

On the Accounts, Users, and Roles screens, record selection for bulk deletion has been fixed. The delete button is unavailable until rows are selected, and the “Select all” action now correctly takes table records into account.

This reduces the risk of accidental deletion and makes bulk operations more predictable.

5. Fixed image context counting#

In requests with images, Sherpa AI Server now counts tokens more accurately before shortening the context and when displaying usage details. Local multimodal models use their own counting mechanism, while cloud models use a single configurable estimate.

No new screen appears in the interface. This change helps avoid unexpected context overflow and image loss in requests to multimodal models.

6. Fixed prompt library folder move#

In the Prompt Library, moving shared and personal folders to another parent folder has been fixed.

This makes prompt organization more reliable and preserves the folder structure chosen by the user.

7. Clarified reranking API behavior#

The POST /api/v1/threads/rerank method now uses the built-in BGE reranker if model_instance_guid is not provided in the request. This behavior is explicitly documented in OpenAPI.

No new screen appears in the interface. This makes integrations simpler: using the default reranker no longer requires selecting a model instance in advance.