Version 2.5.5#

A new Sherpa AI Server release is available: Version 2.5.5.

1. Before upgrading#

Before upgrading, read the notes for all intermediate versions and back up the database, settings, and installation files. If the current version is earlier than 2.5.0, agree on the upgrade path with support.

Existing installations can now switch to external LLM and Embedding services during an upgrade. The previous mode remains the default. Administrators can select external services and check migrations and backups using the upgrade instructions.

2. Model settings for new Chats#

On the Models screen, the system prompt, temperature, and streaming mode are now set for each Model. These settings apply when creating a Chat without an Assistant. A Model can be duplicated with its instances, and its saved settings can be applied separately to related regular Chats. This lets administrators change new Chat settings without changing every existing conversation at once.

In an open Chat without an Assistant, Model selection no longer requires an additional client appearance setting. A Model cannot be changed while a response is running; Chat history remains intact. During an upgrade, the selected legacy Model's settings, including temperature and streaming, are transferred to its record.

3. Message queue and Agent responses#

On the Chat screen, text messages can be queued while an Agent is working. Pending messages appear above the input field and can be viewed or removed. They run in order after the current run or reach the Agent between its steps. This allows a follow-up without stopping the current work.

Step-limit, repeated code interpreter failure, and sandbox errors appear in the interface language. Search tools do not pass chunks without matches to the Model; positive matches are added after reranking up to the configured limit.

4. Chat Moderation#

Administrators can assign a Moderation rule set to all regular Chats in an Account, all Assistants, or selected Assistants. The assignment also applies to new Chats. Rule actions run in sequence: blocking and a fixed response stop further processing. This makes combinations of masking, redirection, and other actions predictable.

Hints and supported action type validation have been corrected in the rule form. A previously saved unsupported rule no longer interrupts message sending; its cause is recorded in the log.

5. Documents and Chat files#

After a document is saved, the Documents list updates without reloading the page. The selected folder, filters, expanded search groups, and list position remain in place. Double-clicking a supported file name opens its preview. See the Documents instructions.

A Chat file link now opens the file in a new tab. XML and other active markup are displayed as text; a separate download action still provides the original file. See the Chat instructions.

A summary of the size and indexing status of nested files is available for Document folders. Administrators enable server calculation with DOCUMENT_FOLDER_SUMMARY_ENABLED=true and display in the interface with show_document_folder_summary in branding.json. Both are off by default. The summary respects access rights to files from private conversations.

6. Home page and prompt library#

Administrators can replace the Home page after sign-in with a single HTML file through client appearance settings. If the file is missing or invalid, the standard Home page opens. An adaptive example with links to application sections is included.

The administrative Account now has a library of 37 shared prompts for documents, reports, and text work. Users with the required rights can select these prompts instead of entering the same requests repeatedly.

7. Roles, Assistants, and Audit#

The Roles list shows associated Users and Object folders. Role details show readable permission names, and a specific User can be opened from the list. This helps verify access before changing a Role.

In the sidebar, Assistants from selected Object folders load after permissions are synchronized. The Audit section localizes event type, message, and filter values when the Russian locale is used.

8. MCP connections#

The MCP card now supports "Bearer token" authorization and a separate tool-call timeout. Leaving the token field empty while editing preserves the existing secret; selecting "No authorization" removes it. Administrators can update a connection without re-entering its token each time.

9. Security and service checks#

Direct access by GUID to other users' private Chats and their messages, files, and runs has been closed. Before files are linked to a conversation, the server checks the entire selected list. An accessible Chat can no longer be used to attach another user's private file.

The delivery now includes health checks for Docker services, including the database, WebSocket, and code interpreter. Upgrades of existing installations also restore a selected legacy Chat Model and refresh database metadata. Administrators can check service readiness after upgrading.