Release 2026.14
4 min read
This page summarizes the 2026.14 release. Click the links below to navigate to the relevant sections.
Chat Plaform | RAG | Agentic Table | UniqueAI |
π Release Overview
2026.14
Unique hosted:
Self hosted:
Chat Platform
π What's New (Business/End-User Focus)
User BETA Code Execution UI Component in Chat: Code execution results are now rendered inline within chat answers as interactive cards, replacing the previous plain-text output. Each card supports multiple output views with a tab toggle between the source code and the rendered result. Supported output types include:
Charts & Images β rendered inline with responsive scaling, aspect ratio preservation, and one-click copy
Tables β CSV/tabular data displayed as a styled, horizontally scrollable table with column headers and alternating row colors
Files & Documents β file download cards with type-specific icons for PDF, DOCX, XLSX, CSV, etc.
HTML is rendered in a sandboxed preview with the option to download as HTML or PNG. It will be unified into the shared code execution component in the coming sprints.
For full documentation please see here:
For Users: Use Unique AI Tools
For Administrators: Code Execution
Admin BETA "Coding Capabilities" section in Sources & Tools β Code Execution has been moved from the general "Tools" subsection into a new dedicated "Coding Capabilities" section within Space Admin β Sources & Tools. These tools do not show up in the tool selection for the end user, but is automatically called by the orchestrator.
For full documentation please see here:
For Administrators: Code Execution
Required Feature Flag: FEATURE_FLAG_ENABLE_CODE_EXECUTION_UN_17498
π§ Improvements
Admin Prompt editor unsaved-changes guard: The prompt editor now shows a confirmation dialog when closing with unsaved changes, preventing accidental loss of in-progress edits to suggested prompts. Empty new prompts that havenβt been typed into no longer trigger a false warning.
π Bug Fixes
Admin Added prefetch limit to the webhook scheduler's AMQP consumer to prevent out-of-memory crashes under message backlog.
RAG
π What's New (Business/End-User Focus)
π Bug Fixes
User Hide failed files from context selection: Files that failed ingestion (e.g. parsing or metadata errors) no longer appear in the βSelect Files to Chatβ context window. This ensures users can only select fully processed files with available content, preventing empty or broken chat experiences.
Agentic Table
π What's New (Business/End-User Focus)
User Insert Row: The agentic table's row context menu supports inserting a new row before or after the selected row.
Admin EXPERIMENTAL Excel (XLSX) Question Extraction: The Agentic Table now supports enhanced question extraction from Excel workbooks, supporting data validations (dropdown lists), hidden option sheets, VML form controls (checkboxes), bold/merged cells, and multi-sheet structure. Enable by setting use_xlsx_preprocessing in the Question Extraction configuration.
Admin New Agentic Table Row Submission Metrics: Two new product metrics β agentic_table_row_submitted_initial and agentic_table_row_submitted β now track the number of rows submitted in Agentic Table sheets. Unlike existing metrics which count per unique user, these metrics increment per submitted row, providing visibility into total question volume and re-run frequency.
π§ Improvements
User Disable file upload for non-owners on shared sheets: The "Import Questions β Upload File" action is now disabled for users who are not the owner of a shared sheet, preventing confusing 403 errors.
User Cell edit mode context menus: When editing a cell, highlighting text now shows a text style menu, while right-clicking shows the native browser context menu. Right-clicking over highlighted text correctly replaces the style menu with the browser menu, consistent across Mac and PC.
Admin Strip references from DOCX export when citations are removed: When the flag strip_citations_from_answers_on_export in the Docx Generator configuration is enabled, the DOCX export now also excludes the References section, producing a clean document with only questions and answers ready for copy-paste.
Unique AI
π What's New (Business/End-User Focus)
Admin Enhanced Execution Time Logging for Agent Workflows: A new execution_time section to the debug info has been added, providing detailed timing breakdowns for each step of the agent workflow. This includes total execution time, per-loop metrics, and granular insights into planning, tool usage, evaluation, and post-processing. These improvements enable better performance analysis and faster identification of bottlenecks. Execution Time Tracking for Unique AI Agent Steps
Admin OpenAI GPT-5.4 Added to Supported Models: Support for OpenAI GPT-5.4 within Unique AI has been added, and it now appears in the model selection dropdown when available to your organisation. Availability is currently limited to global deployments only, with no standard zone deployment in Europe at the moment.
π§ Improvements
Admin Code Execution Tool Usage Now Included in Metrics: A tracking for code execution tool invocations to the agent metrics system has been added, consistent with existing tool logging. This enables visibility in the analytics tool into when and how often code execution is triggered during agent runs.
π Bug Fixes
Admin Respect Custom Hallucination Check Titles in New Chat UI: An issue where the new chat UI ignored customised hallucination evaluation titles and always displayed a default message has been fixed. The UI now correctly reflects the configured titles defined per space (e.g., βAccuracy is Goodβ instead of βNo Hallucination Detectedβ). This restores expected behavior and ensures consistency with client-specific configurations.
Config | Chat |
|---|---|
Admin Preserve Readable Unicode in Internal Search Tool Outputs: An issue has been fixed where non-ASCII content (e.g., Japanese text) in Internal Search tool results was serialized as escaped Unicode sequences before being passed to the LLM (e.g. \u30da\u30fc\u30b8\u540d instead of γγΌγΈε). Tool outputs are now preserved in their original readable form throughout the agent loop and history handoff. This improves log readability, maintains prompt quality, and ensures correct behaviour in multilingual use cases.
Author |
|
|---|