Version 2.4.8#

A new release of Sherpa AI Server is available: Version 2.4.8.

1. Artifacts panel in chat#

A right-side Artifacts panel has been added to the Chat screen. It shows files created during Assistant activity, including files that are not linked separately in the response text. The panel refreshes after the request finishes and also checks for files that appear with a delay.

On mobile devices, the panel does not take space in the conversation area: it can be opened from the Open artifacts menu in the chat header.

This makes it easier to find and download Assistant results right away without reviewing the entire conversation.

2. Updated Chat screen interface#

The central conversation area on the Chat screen has been redesigned: messages are grouped into a compact centered feed with a limited text column width, Markdown readability has been improved, and regular Assistant responses are displayed without a separate card.

For selected text in messages, a floating panel has been added with Copy and quote insert actions for the input field. Secondary actions are grouped under the + menu, the current conversation title is shown in a compact chat panel, and message deletion has been moved to an additional menu. For messages from the current day, only the time is shown without the date. The maximum input field height has been reduced on desktop and mobile devices while preserving internal scrolling.

This makes long responses easier to read, simplifies quote handling, and keeps the input field cleaner on desktop and mobile devices.

3. Chat header and compact Assistant block#

On the Chat screen, the conversation title has been moved back to the left side of the header, and the compact Assistant block is displayed on the right next to the settings button. The block shows files, MCP, packages, and the model; tooltips with detailed information are available on hover.

Detailed Assistant data has been moved from the header into a compact tooltip: it opens when the Assistant name or the info button is clicked and closes when clicking outside the block. Issues with the tooltip overlapping messages, empty space to the left of the chat title, and settings button overflow on narrow screens have been fixed.

In an existing chat, changing or unlinking the Assistant is now available through a modal with tiles like the one on the Assistants screen: search, icon and color, model, pinned at the top, file, folder, MCP, and package counters. The selected chat mode for the current conversation is preserved and restored after navigating between pages.

This makes it easier to understand which Assistant is being used and to switch Assistants safely without losing chat context.

4. Live chat updates and external Assistant response#

Live updates have been fixed on the Chat screen for Assistants that respond through an API: while execution is active, a waiting indicator is shown, and after completion messages are loaded without refreshing the page.

While an external request is in waiting status or another unfinished state, resending and changing the Assistant are blocked in the same way as during internal execution. No extra server-side launch methods are called anymore for an Assistant with an external API response.

This makes integrations more predictable and reduces the risk of duplicate requests.

5. Sending long audio and video#

Message sending with long audio and video has been fixed on the Chat screen: the interface waits up to 10 minutes for transcription and resumes execution if the request is interrupted or the chat is reopened.

This helps complete conversations with long media attachments without manually refreshing the page.

6. Users screen#

On the Users screen, the table now includes columns for user roles and activity status.

This helps administrators quickly assess permissions and account state without opening the user card.

7. Models screen#

In the Create Model and Edit Model dialog on the Models screen, generation parameters have been added: preset, temperature, top_p, top_k, min_p, presence_penalty, repetition_penalty, and sampling_max_tokens. The fields include tooltips in Russian and English.

A chat_template_kwargs setting has also been added for passing additional chat template parameters to external OpenAI-compatible model instances; the value can be set in the model settings or through the LLM_CHAT_TEMPLATE_KWARGS environment variable.

This simplifies fine-tuning generation and integration with external providers without editing the server configuration for each model.

In the Assistant settings on the Widget tab, the Hide “Thinking...” blocks in the widget toggle has been added: when enabled, the reasoning panel is hidden from end users of the widget, while response generation on the server remains unchanged.

On the Assistant edit screen, the system prompt field now has a large resize handle, which is convenient for long instructions and in dark theme.

The moderator setting has been fixed: the selected rule set is now saved correctly, and the No moderation value allows an assigned moderator to be removed.

This helps administrators control what widget users see, edit long prompts more comfortably, and configure response moderation without errors.

The icons for the Assistants, Chat history, Roles, Accounts, and Models sections have been updated in the side menu so that the items remain distinguishable in compact mode without labels.

The portal logo is now a link for creating a new conversation. After sign-in, if no explicit returnUrl is specified, a new chat opens instead of chat history. The Ctrl/Shift selection hint in the side menu is shown only when hovering or focusing inside it.

In the Russian localization, the label was standardized to Assistant name for consistency across the interface.

This simplifies navigation and makes the interface clearer on both desktop and mobile devices.

10. Notifications#

The design of toast notifications has been updated: compact responsive cards with modern typography, a subtle shadow, a status bar, and a close button that is always available. No more than three messages are shown at once; error notifications use a neutral surface with a red accent, and warnings use a yellow palette.

This makes error and warning messages more visible and visually cleaner.

11. Linking files to conversations#

File handling in chat has been fixed: the POST /api/v1/threads/{thread_guid}/files method with file_guids links a file through the thread_files table without changing the source folder. A single file can participate in multiple conversations, and the Assistant knowledge base is not emptied.

Files from file_guids in a user message are automatically linked to the conversation so that agent tools can see the attachment. Clearing or deleting an upload in the conversation folder does not remove the file from disk if it is still linked to other conversations.

This makes it safer to share documents between chats and prevents files from being lost from the Assistant knowledge base.

12. Role-based access to object folders#

When signed in as a user, the list of object folders is restricted to the union of folders available across all of the user’s roles. Signing in as an account still shows all account folders; objects without a folder remain available.

This improves security and matches the assigned role permissions.

13. Agent scenario improvements#

An issue has been fixed where agent scenarios could loop or fail to return a response when the tool schema required thread_guid: the server still injects the current conversation GUID before execution.

The tool selection mode for reasoning models has been updated. If a model reports a tool call but does not return a recognized command, execution ends with a diagnostic error. Agent context compaction has been added: large tool results are limited before being sent to the model, and when the context overflows, the history is compressed with a short progress summary.

Agent request completion after MCP calls has been fixed: the limit for large results is applied only to the copy used by the model, the full history is preserved, and context overflow is shown separately from the general error.

This improves the stability of agent scenarios and reduces the risk of timeouts in long tool chains.

14. Code interpreter#

The Python interpreter no longer executes code from the model response’s internal reasoning blocks; an unclosed reasoning block is also excluded from the executable response.

This prevents accidental execution of reasoning text as Python code.

15. Improved API documentation#

OpenAPI descriptions have been updated for automatic and manual indexing during file uploads, the special mode for transcribable media, and all file status values.

Descriptions have also been clarified for chat creation, model instances, file uploads into folders with object_folder_guid, the multipart force parameter for the interpreter, and API key versus GUI session scenarios.

API documentation is available in Swagger: https://aiserver.sherparpa.ru/api/docs/swagger

This makes integrations easier and reduces validation errors when working through Swagger or curl.

16. API reliability and security#

The GET /api/v1/threads method no longer returns a 500 error if the info or metadata fields in the database are stored in a legacy string format.

It is no longer possible to update or rebind an execution to another account’s conversation through PUT and PATCH requests. Test API routes and their OpenAPI descriptions are available only when application debug mode is enabled.

The file search request from the Designer block, Find embeddings, has been fixed: legacy parameters are converted to the v1 API format.

This improves Sherpa AI Server update stability and the security of installations with multiple accounts.

17. Installation, update, and GPU#

During GPU installation and updates, the gpu compose profile is always enabled so that the local language model service is started, including when Whisper and reranker are not used.

The Nginx API timeout has been increased to 600 seconds in all environments, and the default LLM connection timeout has also been increased. The .env.main template has been updated for the official Qwen3.6 launch in vLLM.

There are no new product screens for these changes: the result is visible to administrators in more stable long requests, correct GPU installation, and updated environment configuration instructions.