📌 Release Overview

• 2026.28

• Unique hosted: July 8, 2026

Chat Platform

🐞 Bug Fixes

Admin Cost Management usability and reliability fixes: Following the initial rollout of Model Usage Limits, a number of issues in Usage Reporting and Limits management have been fixed:

• Table sorting no longer breaks or causes columns to jump.

• Limit values in the Limits view now display their currency unit.

• The "By User" drill-down breadcrumb and the "By Assistant" / "By App" reporting views now resolve to human-readable display names instead of raw internal IDs.

• Links in reporting tables now use a theme-safe text color instead of the primary color (hover underline retained) to avoid clashing with client themes.

• Group limits can now be searched when configuring group-level limits. The redundant "+" button has been removed from the User limits view while keeping its combined filter-and-add functionality.

• Users can now see a clear "Your model usage capacity has been reached" warning when a limit is hit mid-request — including during long-running Conduct tasks — instead of the request silently hanging or failing with a generic error.

RAG

🚀 What’s New (Business/End-User Focus)

User Experimental New Knowledge Base UI refresh: The Knowledge Base has a redesigned browsing experience. Instead of the multi-column (Miller-column) layout, folders and files are now shown in a single sortable list with a side detail pane, making it faster to scan large folders and jump straight to a file's details. Highlights include:

• Sortable list view with sortable columns (name, type, ingestion status, last updated) replacing the stacked column navigation.

• List view filters to quickly narrow down folder contents.

• Detail pane with a redesigned folder header, ingestion status card, and folder analytics shown alongside the list.

• Drag-and-drop upload directly onto the list, with clearer upload and ingestion progress.

Admin Added ingestion metrics: Introduced new metrics that track cumulative file ingestion counts, files added and removed per period, and the upload channel used (Knowledge Base vs. Upload & Ask). These metrics feed directly into the product metrics and also insights pipeline

Admin Embedding model switching now available as a self-service operation: Added the ability to switch a tenant's embedding model (including dimension changes). A new maintenance operation re-creates the vector collection with the correct dimensions and re-embeds all documents, then atomically swaps the live pointer eliminating the need for manual cluster intervention when changing models.

🔧 Improvements

🐞 Bug Fixes

User User Experimental Outlook .msg email content now included in chat context: Fixed an issue where uploading an Outlook .msg file in chat (Upload & Ask) did not add the email body to the model context - the assistant could only respond to the injected citation instructions instead of the actual email. The ingestion pipeline now correctly extracts email bodies stored in compressed RTF format, which is how many real Outlook exports encode message content.

User Security: Fixed SQL injection vulnerabilities in the RAG search layer (similarity search, full-text search, and access-control filters). All user-supplied inputs are now parameterized.

User Citation now opens the correct page in PDF preview: Fixed an issue where clicking a citation in the chat response would navigate to the wrong page in the PDF viewer, even though the correct chunk was retrieved from the knowledge base. The page mapping logic between chunk metadata and the PDF renderer has been corrected, so citations now scroll to and highlight the exact source passage.

User Word DOCX date field picker now reads the correct date: Fixed an issue where date values selected via dropdown date pickers in Word documents were incorrectly read as 30 December 1989 during ingestion. The pipeline now correctly parses structured date fields from DOCX form controls, preserving the actual selected date.

Agentic Table

🚀 What’s New (Business/End-User Focus)

User GA Agentic Table Comments now generally available: The comments feature has been promoted to GA and is now enabled for every sheet, no feature flag required.

🔧 Improvements

User Text highlighting in agentic table cells: Text inside cells can now be color-highlighted using the <mark> tag. The highlight uses the configurable attention theme color, so it adapts to each theme. Excel export preserves the text but drops the highlight.

User Preconfigured metadata keys protected from manual override: When a metadata key is already preconfigured via the sheet metadata form, manually adding the same key is now blocked with a clear message pointing to the form, instead of silently overriding the value.

User Section and Question columns locked for “can answer” users: Users with the "can answer" role can no longer edit the Section and Question columns. The question structure stays protected while editors and owners maintain full access privileges.

User “Can answer” users can no longer remove themselves: A “Can answer” user cannot remove themselves from the assignee list and may only remove assignees they added themselves; owners and editor privileges are unaffected.

🐞 Bug Fixes

User External link protection in sheet cells: Links in sheet cells no longer redirect users silently. Clicking an external link now opens a confirmation modal that shows the full destination URL, letting the user decide whether to proceed — closing an open-redirect vector where malicious links could be injected into shared sheets.

User References side panel now shows the correct sources for Subagent Answers: In the agentic table, the references side panel could display sources that didn't support the Subagent Answer. When library sheet metadata was unavailable, the fallback attached the table row content itself as the reference instead of the library file's content chunks. The fallback now correctly uses the library file content chunks, so the references shown in the side panel actually substantiate the answer.

Language Models

Admin Claude Sonnet 5

Sonnet 5 is now available through LiteLLM. Two deployments mode are possible:

• litellm:anthropic-claude-sonnet-5: Directly through anthropic API

• litellm:vertex-claude-sonnet-5: Through VertexAI (processing restriction to EU)

Admin GLM 5.1 & 5.2

GLM 5.1 and GLM 5.2 are now available through LiteLLM (hosted via Together.ai).

• litellm:glm-5.1

• litellm:glm-5.2

Admin Model retirement

The following models are deprecated and not further available after July 10, 2026:

• litellm:qwen-3-235B-A22B

• litellm:qwen-3-235B-A22B-thinking

• litellm:glm-5.1

• litellm:deepseek-r1

• litellm:deepseek-v3-1