Version 2.4.6#

New release of Sherpa AI Server: Version 2.4.6.

1. Mobile and tablet interface adaptation#

The Sherpa AI Server interface has been adapted for smartphone and tablet screens. The side menu now opens as a slide-out panel, dialogs, notifications, and the Chat screen adjust to the available space, and table rows can be viewed as compact cards with key properties and expandable additional data.

On screens with tables, a toggle between cards and table view and a compact actions menu have been added. The Models and Moderation screens also now have dedicated mobile layouts. Control sizes, tap targets, message spacing, input fields, menu behavior, and assistant cards have been improved.

This makes the core Sherpa AI Server features more convenient on mobile devices without changing the familiar desktop layout.

2. Shared and private chats#

The Chat screen now includes shared and private chats. The chat creator can change its availability and publish a link of the form /shared/thread/{guid}. Through the public link, the chat opens in read-only mode, even without authentication, and is not indexed by search engines.

For authenticated users, the standard chat list and pinned items remain available. After permissions are synchronized, shared chats and assistants load correctly in the side menu, and when signed in under an account, all chats associated with it are displayed, including private ones.

This helps safely separate personal work and materials for collaborative viewing.

3. More transparent access management#

The role creation and editing dialog has been updated on the Roles screen. On the Access tab, an adaptive dependency table and a separate “Menu folders” list are shown next to the “Accesses / Rights” table. They show which read permissions are required for menu sections and which sections the role cannot access.

Unavailable side menu items are displayed in gray.

After a role is changed or deleted, permissions and the menu are updated without reloading the page, and navigation to an unavailable page redirects the user correctly. Access to Object Folders is also taken into account for the Documents and Models screens.

This simplifies role setup and helps administrators see in advance how menu sections depend on assigned permissions.

4. More stable document search and reranking#

BGE reranking has been improved: the number of candidates and the query batch size are limited by settings, long documents are evaluated in several overlapping fragments, empty fragments are not sent for processing, and after unavailable files are filtered out the system fetches the required number of results.

If BGE is unavailable, a warning is shown on the Chat screen and the standard search order is used. Enabling reranking for an assistant now also applies to previously created chats.

The responses of POST /api/v1/files/search and POST /api/v1/files/searchbm25 now include the rerank_applied, rerank_skip_reason, vector_distance, bm25_score, rerank_score, and rerank_score_normalized fields. The score field is preserved for backward compatibility.

This makes document search more robust, especially when working with long files, and helps administrators and integrations understand how the result order was formed.

5. API improvements for external integrations#

The Assistant external response mode has been fixed. The execution remains in WAITING status without calling the internal language model and is completed by the external API client with the CHAT_COMPLETED status.

The text_chunk field, which is actually returned, has been added to the OpenAPI description of the file search response. Shared and private chats are separated by user, and the filtering and visibility parameters have been aligned with the current API contract.

This simplifies connecting external systems and reduces the risk of errors caused by discrepancies between API behavior and its documentation.

6. Authentication and moderation fixes#

On the sign-in page, server errors are now displayed as a standard notification instead of an HTML response under the input fields. Protected pages no longer show their content before the session is successfully checked, and missing permission for a section is handled correctly before it is displayed.

Dictionary trigger checks for moderation are now case-insensitive for UTF-8, including Cyrillic, both when testing rules and during message processing.

This makes error messages clearer, prevents brief exposure of unavailable content, and improves moderation rule accuracy.

7. New settings for administrators#

The favicon_url parameter has been added to branding.json, allowing the browser tab icon to be replaced. For the GPU LLM server, the LLM_EXTRA_ARGS variable allows additional vLLM arguments to be passed after the main startup parameters.

Docker Compose service configuration is now loaded from .env without duplicating most variables in environment. These changes do not add new screens to the interface.

This simplifies branding, vLLM configuration, and environment configuration maintenance.