📌 Release Overview

• 2026.24

• Unique hosted: June 10, 2026

Chat Platform

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

User Experimental Switch the AI model in a chat: You can now choose which AI model answers your message directly from the chat input, picking from a list of models your space admin has approved. Match the model to the task — a faster model for quick questions, a more capable model for complex reasoning — without leaving the conversation or starting a new chat.

Admin Experimental The models users can switch between can now be curated: In a space's Configuration, enable "Allow users to change the model" and select the additional language models that appear in the chat model dropdown. Models are grouped by provider and the space's default model is always included and can't be removed. Custom model deployments can be added from the same place.

User Experimental Expand code execution results in a side panel: Code execution results — charts, tables, rendered HTML, documents, and source code — can now be opened in an expanded side panel. Click the Expand button on a code execution card and the panel slides in from the right (~50% of the screen), giving a larger, more readable view of the output. The panel header includes a code/output toggle, copy and download actions, and closes via the X button or by clicking outside. This builds on the inline code execution card introduced earlier.

Supported file types in the document view: PDF, PowerPoint (.pptx), Word (.docx), PNG, and CSV. PowerPoint and Word files are converted to PDFs for viewing.

🔧 Improvements

User JSON files can now be successfully uploaded: The issue with uploading JSON files to a chat has now been resolved.

User Chats created from uploaded files get their title updated on the first prompt: Previously, when a chat was started from uploading a file to chat, the chat was titled by the file name, which sometimes would not make much sense; now the chat title gets updated on sending a first prompt, which gives it a more meaningful and closely related title based on the prompt.

Admin MCP Connectors are searchable: MCP connectors from MCP hub are now searchable and appear lexicographically sorted. MCP servers still remain listed first then virtual servers and are sorted independently.

User Open a space's access settings from the sidebar: The space options menu in the sidebar now has an "Access" action that opens the sharing dialog for that space directly — so you can add or remove members, change someone's access level (View / Manage), and copy the space link without opening the space first.

Admin Chat references style: The chat references style has now become a company configuration and is no longer handled via feature flags. The existing feature flags are overriden and configuration has been automatically migrated to the company’s configuration.

Admin Feedback modal redesigned with sensitive content consent: The feedback modal has been redesigned to match the latest design specs. Message content (space name, user prompt, and agent response) is now always included in feedback CSV exports. An optional sensitive content consent checkbox can be enabled per company, letting users attest before sensitive content is shared. The CSV export surfaces the new includeSensitiveContent column.

Admin Persistent banner moved to Advanced Settings: The persistent notification banner configuration has been relocated to Advanced Settings in the admin panel. The banner now renders below the header (outside the scrollable chat area) and its styles have been updated to match Figma design tokens across all variants (information, warning).

User Dropdown sub-menus now scroll: Dropdown sub-menus with many items now scroll instead of overflowing past the viewport edge.

Admin New space access modal enabled by default: The new space access sharing dialog (previously behind FEATURE_FLAG_ENABLE_NEW_SPACE_ACCESS_MODAL_UN_17544) is now enabled by default. The feature flag has been removed.

🐞 Bug Fixes

User Elicitation form multiline fields now render correctly: Fixed the "multiline": true field option being ignored in elicitation forms. Fields configured as multiline were still shown as single-line inputs. They now render as a multiline text area as intended. A field also renders as a multiline text area when the schema implies long text (a multi-line or long default value, a high minLength, or a large maxLength). Short fields and enum/email/URL/date fields stay single-line.

Admin Notification banner dates: Fixed the dates not properly reflected in the UI after saving them.

User Feedback modal width consistency: Fixed the feedback modal shrinking when fewer elements are rendered (e.g. positive feedback without consent). The modal now maintains a consistent 558px width across all feedback flows, including the translation page.

User File names preserved on clipboard paste: Pasting images or files into the chat input now keeps their original file names instead of appending numeric suffixes (e.g. image.png instead of image1.png).

RAG

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

User Exclusive folder navigation: Users with file-level access can now reach Knowledge Base sub-folders they own or are explicitly authorized for, even when they do not have access to ancestor folders in the path. Parent folders that appear only because of exclusive descendant access are shown with a configurable label (default [Restricted]) instead of their real name, so users can navigate without exposing sensitive folder names.

Note: Requires both FEATURE_FLAG_ENABLE_FILE_BASED_ACCESS_UN_12660 and FEATURE_FLAG_ENABLE_FOLDER_PATH_RESOLUTION_FOR_EXCLUSIVE_FILE_ACCESS to be enabled for the company. Applies consistently in the KB tree, Select Files to Chat picker, and internal search.

🔧 Improvements

Admin Configurable label for restricted parent folders: A new node-ingestion setting EXCLUSIVE_FILE_ACCESS_FOLDER_LABEL (default [Restricted]) controls the display name for parent folders visible only via exclusive descendant folder access. This has been set to an empty string to show real folder names instead.

Admin User Handling of dynamic rules for internal knowledge in preview pane in space admin and chat: The preview pane no longer applies the dynamic rules in the preview.

🐞 Bug Fixes

User Delete own chat uploads without Knowledge Base write access: Users who upload files to chat in the New Answers UI can now remove their own uploads even when they do not have chat.knowledge.write or chat.kb.read permissions. Previously, the delete action returned a 403 Forbidden error.

User Folder navigation 403 errors: Fixed intermittent 403 Forbidden errors when browsing between Knowledge Base folders, caused by incorrect scope-management queries for the current user's group memberships and grants.

User Word and PowerPoint ingestion on clusters without PDF preview: Fixed Word and PowerPoint documents ingesting with zero chunks on clusters where CONVERT_OFFICE_FILES_TO_PDF was disabled. After the default Word ingestion mode switched to MS Document Intelligence (ingest-as-PDF), those files depended on Office-to-PDF conversion but were incorrectly blocked by the preview flag. Ingestion now always converts for the ingest-as-PDF paths; chat preview behavior is unchanged.

Admin Move Folder dialog navigation: Fixed the Move Folder dialog hiding ancestor folders that the user cannot write to directly, even when they contain writable nested children. Non-writable folders are now shown as navigation nodes (but remain non-selectable as destinations), so users can expand through them to reach writable targets.

Agentic Ingestion

🔧 Improvements

Admin Remote Feature Flag Evaluation: The three agentic ingestion feature flags are now evaluated per-request via backend-service-configuration, enabling per-company and per-user toggling from the admin UI without service redeployment. The flags are registered under their full names

• FEATURE_FLAG_ENABLE_AGENTIC_METADATA_EXTRACTION_UN_15619 (default: false)

• FEATURE_FLAG_ENABLE_IMAGE_CONTENT_EXTRACTION_UN_17223 (default: false)

Note: Self-hosted customers must set CONFIGURATION_BACKEND_URL on ai-service-agentic-ingestion to activate remote evaluation (see Infrastructure Changes). Without it, the service fails to run.

Agentic Table

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

User Copy answer from table cells: Right-click an answer cell to copy a clean version of the answer to the clipboard, with citation references and supporting notes removed. This feature is behind a feature flag called FEATURE_FLAG_AGENTIC_TABLE_COPY_ACTION_UN_20824 and must be enabled per environment before it is available.

🔧 Improvements

User Long table names are truncated: Long Agentic Table names are now truncated so they no longer stretch the table off-screen. Hover over the name to see the full title in a tooltip.

Admin Configurable subagent timeout per column: Each subagent column can now set its own response wait timeout (default 300s, previously a fixed 120s), preventing premature timeout errors on longer-running answers. See the "Columns" section for details.

Admin Cleaner Excel exports (markdown & HTML): Agentic Table Excel exports can now strip HTML tags and render markdown as native Excel formatting (bold, italics, lists, etc.) instead of showing raw syntax and <details> blocks. Both are opt-in and off by default; see the "Xlsx Generator" section for the exact configuration.

Admin Per-column search scopes for subagent columns: Subagent (Agent) columns in the Agentic Table can now be restricted to specific document scopes, so each column answers only from the sources you intend. Two controls are available:

• Search Scope ID (scope_ids) — pins a column's retrieval to a designated scope (e.g. a curated answer library or a specific folder). The column searches only that scope and is shielded from other content — including chat-uploaded question files — preventing cross-source contamination. For example, a "Library Answers" column can be configured to draw exclusively from the answer library, while a fund column stays restricted to that fund's documents.

• Additional Scope IDs (additional_scope_ids) — extends a column's retrieval to one or more extra scopes on top of its primary scope, broadening coverage when an answer may legitimately live across several folders, without losing the scoping guarantees above.

Together these give admins precise, predictable control over what each column can see, improving answer accuracy and isolation in multi-agent sheets.

🐞 Bug Fixes

User Due date update access control: Fixed a security issue where unauthorized users could update another user’s sheet due date if they knew the sheet ID; the mutation now verifies the caller has proper access.

User Improved reliability when re-running a sheet: Fixed a broken search query that could cause the subagent to return empty responses on sheet re-runs, addressing a likely cause of the reported reliability issue.

User Assignee column respects space membership: The Assignee column no longer lists users who have been removed from the space. Only current space members can be assigned, matching the Share dialog.

User Agentic Table files excluded from retention cleanup: The FILE_RETENTION_IN_DAYS cleanup now applies to chat files only. Agentic Table files are no longer deleted by the retention window.

Unique AI

Language Model

Admin Claude Opus 4.8

Opus 4.8 is now available through LiteLLM. Two deployments mode are possible:

• litellm:anthropic-claude-opus-4-8: Directly through anthropic API

• litellm:vertex-claude-opus-4-8: Through VertexAI (processing restriction to EU)