Release 2026.16

6 min read

📌 Release Overview

• 2026.16

• Unique hosted: April 15, 2026

• Self hosted: April 23, 2026

Chat Platform

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

User PowerPoint File Support: Chat now correctly recognizes and renders uploaded PowerPoint (.pptx) files with the appropriate file type icon and preview behavior.

Admin MCP Tool Settings Refresh: Administrators can now refresh tool settings for connected MCP servers directly from the Admin Panel. This allows admins to pull the latest tool definitions from an upstream MCP server without needing to disconnect and reconnect.

How it works:

On the MCP server detail page, click Refresh Tool Settings in the tools section toolbar. A confirmation dialog explains that the refresh will replace existing tool settings with the latest upstream configuration. After confirming, the platform reconnects to the MCP server, fetches the current tool list, and reconciles it against the stored configuration.

After a refresh, the Admin Panel shows:

• A summary of what changed: how many tools were added, updated, or removed

• Per-tool badges indicating whether a tool is New, Updated, or Unavailable

• A Last updated timestamp so admins know when tools were last synced

Key behaviors:

• New tools discovered upstream are added as disabled by default, giving admins explicit control before they become available to users

• Updated tools whose input schema has changed are automatically disabled so admins can review the changes before re-enabling

• Removed tools that are no longer provided by the upstream server are marked as unavailable and disabled

• If the MCP server requires OAuth, the refresh flow handles re-authentication automatically before fetching tools

Note: This feature is behind theFEATURE_FLAG_ENABLE_MCP_TOOL_REFRESH_UN_19255 feature flag.

Admin Model Usage and Cost Management (Experimental) - gives organisations visibility into AI model consumption — tracking token usage and estimated spend per user, assistant, and application.

Key highlights:

• Records every LLM call — input tokens, completion tokens, and estimated cost in USD

• New Cost Management section in Admin Settings with usage reporting and model pricing table

• CSV export available for offline analysis, with privacy settings (pseudonymisation/anonymisation) respected

• Costs calculated as: (inputTokens × inputCostPer1M + completionTokens × completionCostPer1M) / 1,000,000

• Default pricing sourced from Azure OpenAI and LiteLLM — per-client price overrides available on request

• Usage can be filtered by date range, and broken down by model, user, or assistant

Note : Controlled by two feature flags, both disabled by default — contact Unique to enable:

• FEATURE_FLAG_SAVE_MODEL_USAGE_UN_12832 — starts recording usage data to the database. Must be enabled first.

• FEATURE_FLAG_SAVE_MODEL_USAGE_DASHBOARD_UN_18889 — enables the Cost Management section in Admin and the spend badge in Chat

Required roles:CHAT_FEEDBACK_READ or CHAT_DATA_ADMIN

Full Documentation: Model Usage and Cost Management

🔧 Improvements

User HTML Preview Height and Error Handling: HTML previews in code execution results now manage their own height correctly, display a loading state while content renders, and show clear error messages when the preview fails to load.

User Hide Suggested Follow Up Questions: For users who prefer a more focused chat experience, suggested follow-up questions can now be hidden with a single click. The preference is saved automatically and persists across sessions.

🐞 Bug Fixes

User Iframe Height Resize Loop: Fixed an issue where embedded HTML content (e.g. Plotly charts with height:100%) could enter an infinite resize cycle, causing the iframe to grow endlessly.

User Advanced Formatting Renders During Streaming: File, image, and HTML previews in assistant responses now appear progressively while the message streams instead of waiting until the response completes.

User Tool Selection Reset on Shared Chat Links: Opening an existing chat via a shared URL no longer resets the selected tools and files to defaults.

User Iframe Height Not Shrinking on Filter: Filtering rows in an embedded HTML DataTable now correctly shrinks the iframe height to match the reduced content, eliminating blank gaps below the table.

RAG

🐞 Bug Fixes

User CSV Upload: Improved how CSV file uploads are handled to make the experience more consistent across different operating systems. CSV files created on Windows will now be uploaded and ingested the same way as those created on another operating system.

User Content Permissions Hardening: Permission checks for file metadata updates are now correctly enforced, ensuring users can only modify files they have access to.

User Scope Permissions Hardening: Permission enforcement for access management has been strengthened, ensuring only users with the "Manage" role can modify access settings.

Agentic Ingestion

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

AdminUser Agentic Image Content Extraction (Experimental): PDF documents often contain charts, diagrams, and infographics whose content is invisible to traditional text extraction. With Agentic Image Content Extraction, the platform now detects individual figures on each PDF page using Microsoft Document Intelligence, crops them, and sends them to a vision-capable language model for AI-powered content extraction. The extracted descriptions are merged back into the page at the correct position, making previously locked visual information searchable and available to the AI assistant.

Importantly, this process enriches the document with textual descriptions of visual content to improve retrieval. The original images themselves are not passed to the agent at runtime. Instead, when relevant content is retrieved, users can view the original figures (e.g. from PDFs, Word documents, or PowerPoint files), while the agent operates on the extracted text representation.

Image Content Extraction can be enabled through two independent paths:

• Knowledge Base uploads — configure via the folder-level ingestion settings in the Knowledge Upload app (toggle + language model selection)

• Chat uploads — configure via the assistant’s advanced ingestion settings in the Admin app

Note:

• This feature requires Microsoft Document Intelligence, a vision-capable language model with structured output support (e.g. AZURE_GPT_4o_2024_1120) and the deployment of the new service agentic-ingestion (in case that it was not deployed yet). Please reach out to us if you would like to test this new feature.

🔧 Improvements

Admin Metadata Schema Validation (Agentic Metadata Extraction): Extracted metadata is now validated field-by-field against the configured JSON schema before being persisted. Each field is checked for type correctness (string, number, boolean, array). Fields that fail validation are silently dropped, while fields that pass are merged into the document’s metadata. If no fields pass validation, the document’s existing metadata remains unchanged. This prevents malformed or hallucinated LLM output from corrupting stored metadata.

Admin Metadata Schema Enum Support (Agentic Metadata Extraction): Metadata field schemas now support an enum property to constrain extracted values to a predefined set of allowed options. When a field defines an enum array, the LLM is instructed to choose only from those values via structured output, and the result is validated server-side: scalar fields with a value outside the allowed set are dropped, and array fields are filtered to keep only matching entries. This is useful for fields like document category, language, or status where only specific values are meaningful.

Agentic Table

🔧 Improvements

Admin Configurable cap on extracted rows: Agentic Table can limit how many rows/questions are extracted from a file (default 500) to protect the cluster from very large extractions. When the limit applies, users get a toast that states how many rows were detected and how many will be processed. More information can be found in the Question Extraction section.

User File names truncated in files dropdown: Long file names no longer stretch the dropdown; names are truncated for a stable layout.

User Knowledge Base modal loading state: The Knowledge Base modal shows a skeleton loader on first open instead of an empty white panel while content loads.

🐞 Bug Fixes

User Row delete on shared sheets: When a sheet is shared with edit access, collaborators can remove rows reliably. Previously, delete sometimes failed despite edit access; that case is fixed.

User Sheet status after deleting an empty row: Deleting a newly added empty row on a completed sheet no longer incorrectly moves the sheet to “Ready for Completion”.

Unique AI

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

Admin Space 2.0 Sub-Agent Configuration: Space 2.0 now supports sub-agents in UniqueAI, allowing admins to define reusable sub-agent spaces and attach them to orchestrators. This lets one assistant space delegate focused tasks to another pre-configured space, so specialized workflows can be exposed as callable sub-agents. When an orchestrator uses one, it sends the request to the linked sub-agent, waits for the result, and continues with the returned content. Setup and usage details are available here.