Version 2.4.10#

A new release of Sherpa AI Server is available: Version 2.4.10.

1. Transcription and OCR#

Now AI Server can process large audio files for transcription and images for OCR without hitting the model context size limit, even if the full transcript or the text recognized from the image does not fit within the model context.

This helps the Assistant work with long transcripts without overloading the request with the full text and understand which file the transcription belongs to.

2. Table and presentation indexing#

XLSX table indexing now takes structure into account: separate subtables, headings, and rows with column context. The fragments preserve the table context and do not split rows.

For legacy .ppt presentations, LibreOffice Impress has been added to the indexing service Docker image. On the Documents screen, when indexing fails, the file status shows the reason. If a presentation cannot be indexed, the status and the chat search response indicate that the file is corrupted or is not a valid PowerPoint file.

This improves search quality for tables and presentations and makes indexing errors easier to understand.

3. Prompt library#

The "Prompt Preview" dialog has been updated: structure and responsive layout, text in a separate scrollable area, and a "Copy text" button.

This simplifies working with long prompts before inserting them into a message.

4. Chat screen reliability#

On the Chat screen, when switching to another dialogue while a message is being sent, stale operations are canceled: file loading and indexing, start and resume execution, and incremental response streaming. The message no longer goes to the wrong dialogue.

Assistant response sources are grouped by file: multiple fragments from a single document are no longer shown as separate sources.

This helps preserve context when switching dialogues quickly and makes response sources easier to read.

5. Chat modes and the "New dialogue" button#

On the Chat screen, the modes dropdown is now wider than the selector button. Long items such as "Full-text search" and "Code interpreter" are fully readable and do not overflow the panel.

The "New dialogue" button in the chat side menu no longer shows a hover tooltip. In the additional actions menu, the spacing between the icon and the text of the "Prompt library" item has been fixed.

This makes mode selection easier and the chat menu items look neater.

On the sign-in form, when credentials are invalid, localized "Invalid login" and "Invalid password" messages are shown below the fields instead of the API's English text.

This helps identify which field needs to be corrected more quickly.

7. Chat history export#

On the Chat History screen, XLSX export no longer fails on messages containing document fragments.

This makes it possible to export the full dialog history, including responses that reference documents.

8. Roles screen#

On the Roles screen, the role delete button now sends the delete request after confirmation and refreshes the "Roles" table. Previously, the role was not deleted after confirmation.

This restores the expected role deletion behavior.

9. Chat error messages#

On the Chat screen, common model request execution errors are shown in Russian and English. The server now sends a stable error code; for older messages, the previous error text is preserved.

This helps explain the cause of the failure and what to do next: retry the request, switch the model, or shorten the context.

10. RTF files#

For RTF files, words no longer run together: text search now finds individual words. Downloading .rtf files from the chat and Documents screens no longer fails.

This restores search and download support for RTF documents.