Release 2025.34

This is a preview page for the upcoming 2025.34 release to announce the necessary infrastructure changes early before the release as can be seen in 2025.34 Infrastructure Changes.

📌 Release Overview

• v 2025.34-de229

• Unique hosted: August 21, 2025

• Self hosted: August 29, 2025
  

👥 Role-Based Highlights

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

User Agentic Table: Implement single-page sheet creation flow: Sheet creation in Agentic Table spaces is now streamlined into a single page, combining name, questions, and sources in one unified view. The feature requires the FEATURE_FLAG_AGENTIC_TABLE_ONBOARDING_V2_UN_12041 to be activated for the CHAT frontend application.

User Add due date support to Agentic Table sheets for improved tracking and planning: Introduced due date functionality across sheet creation, overview, and detail views to help users better manage and prioritize their RFP workflows.

User Admin Experimental ⚠ Model Context Protocol (MCP) support: Adds experimental support for connecting Unique to remote MCP servers. Admins can add MCP servers and exercise fine-grained control over which tools are allowed. They can also add custom user and system prompts to define how the MCP server is used. Additionally, Space admins can whitelist and approve MCP tools in spaces, making them available to users.
MCP tools added to a space are available for users to select, enforce, or set as a default to help them with their tasks or questions.
The feature requires the FEATURE_FLAG_ENABLE_MCP_UN_11963 to be activated for our ADMIN and CHAT frontend applications.
The node-chat backend requires three new mandatory environment variables:

• CACHE_REDIS_DB: A Redis DB number between 0 and 10 to be used as a cache backend for various tasks in node-chat. The DB can be shared with other Redis tasks as all cache keys are namespaced, but ideally a standalone DB is used.

• CHAT_FRONTEND_URL: The URL of the CHAT frontend. Used to create an oAuth redirectUrl so third-party MCP servers can redirect back to Unique during an oAuth flow for users.

• ADMIN_FRONTEND_URL: The URL of the ADMIN frontend. Used to create an oAuth redirectUrl so third-party MCP servers can redirect back to Unique during an oAuth flow for admins.

User Admin (Beta) Agent-to-Agent communication in Unique AI: Unique AI can now be setup with tools that run user requests in separate spaces and return the result to the main Unique AI Agent. Use cases:

• Re-use well tested/optimized spaces.

• Reduce complexity of tasks the main agent has to perform by allowing it to use specialized sub-agents.

We call this tool SubAgentTool . Details, setup and configuration instructions can be found here.

User GPT5 model family: Unique AI now supports all gpt-5 models. Gpt-5 will by default use minimal reasoning effort. Reasoning effort can be overridden using JSON config

• gpt-5 (2025-08-07)

• gpt-5-mini (2025-08-07)

• gpt-5-nano (2025-08-07)

• gpt-5-chat (2025-08-07) [Only supports GlobalStandard]

User File level metadata editing: Added metadata viewing and management in the Knowledge Base file side panel. Users can easily add, edit, or delete metadata. This update enhances metadata visibility and management.

🔧 Improvements

Admin Share functionality: When the user clicks "Share" in the new Space Management, the content is immediately shared without requiring any additional taps or clicks.

User Chats view: In Space management app for “Active” and "Last Edited" chats stay on a single line. If no active chats are available, the chip is hidden.

Admin Translator V2: You can now save each individual LLM call made during document translation. These messages are saved onto the chat message, but are not visible in the UI and are intended solely for debugging or compliance purposes. To enable this feature, set the parameter saveTranslationCallsEnabled to true (see here for config).

Admin Translator V2: By default, few-shot learning examples are included in LLM calls for both text and document translation. You can disable this behavior by setting the parameter translationExamplesIgnored to true, which ensures that no examples are added to the LLM call. See here for the configuration.

User Improve Agentic Table performance by reducing redundant internal service calls: Added caching to optimize internal service requests during high-frequency agent updates, reducing load and improving responsiveness.

User Change default scope in Agentic Table to exclude source documents when none are selected: Updated default behaviour so that when no source documents are selected, the model used by the agent runs without internal search access and relies solely on its built-in knowledge.

 User Align Agentic Table search with configuration and internal search logic: Align Agentic Table’s internal search parameters with Chat’s internal search to provide users with a consistent and comprehensive experience matching that of Unique AI chats.

Admin Unique AI - Upload and Chat: The allocation of tokens for uploaded content has changed from a fixed number to a percentage of the model’s input tokens. approximate_max_tokens_for_uploaded_content_stuff_context_window has been renamed to percent_for_uploaded_content in the Uploaded Content Config (see here for config)

User Unique AI - Uploaded content token estimation improvement: The logic for estimating tokens in uploaded content has been refined to be more accurate, allowing users to fit more tokens into the model than before.

User Unique AI - Maximum tool calls per iteration: The default value of the maximum number of tool calls per iteration has been updated from 10 to 5 for performance optimisation.

Admin Internal Search Tool: The internal search tool has a new configuration parameter Score Threshold that filters the search results based on their similarity score to the (see here for config).

tech Backend Chat: Added possiblity to configure timeout duration for LLM calls. Useful for self-hosted thinking models that need more time to initial response.

User Select Files to chat:

1. The "Select Files" button is now always visible when supported, making it easier for you to access this feature.

2. You can now configure the "Select Files to chat" option in Space management, enabling this functionality for custom modules as well. To enable this functionality, set isSelectFilesToChatEnabled to true for the tools in the custom module.

User Merge Parent with Sub‑folder Metadata Tags: Improved metadata propagation for folder ingestion: parent metadata tags with “propagate to sub-folders” enabled now merge with, rather than override, sub-folder metadata tags.

User Programmatically Handle HTML Tables Instead of converting extracted html tables (e.g., by MDI) to markdown before chunking the extracted text, we leverage the html representation to split tables across Chunks more accurately considering header and merged cell information.

User Improved Folder & Content Loading with Pagination: Implemented pagination for both folders and content, with folders loaded first to improve navigation and maintain performance when many folders exist in a scope.

🐞 Bug Fixes

User Delete several failed files in KB: Issue resolved, users can now delete multiple failed files simultaneously.

User Unique AI - In-chat uploaded content excluded from internal search results: Issue resolved with content uploaded to a chat being visible to the model both in internal search results originating from the knowledge base and as uploaded content.

User Unique AI - Fails when temperature param is passed: Issue resolved with clamping min/max temperature for each of reasoning models

User Table headers should be included in each chunk: Tables are now split row by row until the defined chunk size is reached, with headers at the top of each chunk.

User Excel ingestion skips last row of tables: Resolved the issue where the last row of a table was skipped if the table extended to the final row of the sheet, preventing data loss and ensuring correct chunking.

User validAsOf not editable: Fixed the issue that the automatically extracted validAsOf date would always overwrite the value set through manual user input. Now this value can be overwritten.

.