This is a preview page for the upcoming 2025.50 release to announce the necessary infrastructure changes early. This is our last release for this year. The next release will be on Wednesday January 7, 2026.

📌 Release Overview

• 2025.50

• Unique hosted: December 10, 2025

• Self hosted: December 31, 2025 (delay due to release break over holidays)

Chat Platform

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

User Improved Default Code Execution: A new system prompt will deliver more consistent results when returning images files, and markdown text from the Code Execution Tool.

Admin Targeted App Event Distribution: Users can now select which AI modules should receive events when creating or editing an endpoint for an app in the Apps section, so that events are only sent to the intended modules and not to all subscribed apps.

RAG

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

User Ingestion size limits for Excel and CSV files

We have introduced ingestion size limits for Excel and CSV files. Previously, no limits existed, meaning any file size or dimension could be uploaded, which created potential performance and stability issues. Large excel / csv files could block the ingestion pipeline for other users without guaranteeing a successful ingestion. These files remain accessible for agents by download, e.g., in order to use them via code execution.

With this release, ingestion limits are now enforced — and fully configurable via API (see here).

Default Limits

All scopes now have system defaults. These defaults apply automatically unless overridden:

{
  "excelConfig": { "maxCols": 100, "maxRows": 5000 },
  "csvConfig":   { "maxCols": 100, "maxRows": 5000 }
}

Admin Automated Ingestion Timeout Detection & Recovery

Files can get stuck in certain processing states. We implemented a maintenance timeout service that automatically detects and marks ingestion processes stuck in processing states (reading, chunking, embedding, etc.) for more than 2 hours. The ingestion state for these files will be set in the frontend to Ingestion failed (Unknown Error occured - File stuck in processing state).

Please check the Infrastructure Changes for more details on how to configure it.

🐞 Bug Fixes

User Security Fix for Unlimited Folders Creation in Other Users Private Knowledge Base

A security researcher of our bug bounty program has discovered a broken access issue which allowed the attacker to create new directories in other users' knowledge base and the user cannot delete these directories as the resource ownership was still with the attacker. We closed this security gap and this action is not possible anymore.

User Security Fix for Improper Access Control - Unauthorized Metadata Creation

A security researcher of our bug bounty program discovered a broken access control which allowed the attacker to update/add meta data on other users' files. This worked for files under both public and private directories. For successful exploitation the attacker would have needed to get multiple complex IDs and the researcher did not shown a way to leak these IDs.

To close this security gap, we have introduced authorization controls for content metadata updates to prevent unauthorized modifications. Users can now only update metadata for:

• Content they personally own, OR

• Content in scopes/directories where they have write access

User Security Fix for Unauthorized Read Access and Enumeration of Private User Folders

A security researcher of our bug bounty program discovered a broken access control issue which allowed the attacker to query other user’s folder names with creation and update dates from the knowledge-upload feature of the application. We have introduced access control validation for folder operations to prevent this unauthorized read access.

Users without proper permissions receive a ForbiddenException:

Access denied: You do not have permission to access folder {folderId}

User Improved Display of Metadata in Chunk View

This update fixes missing metadata in the Chunk view, providing a more complete and consistent display.

User Fix of issues ingesting large files using Docling

We fixed a concurrency issue in our Docling-based PDF ingestion pipeline that could cause large documents to fail during processing. The root cause was a single shared DocumentConverter instance being used across multiple worker threads, which is not thread-safe. We now create a thread-local Docling converter per worker, ensuring isolated model state per thread and significantly improving reliability and stability when ingesting large PDFs.

Agentic Table

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

User User Feedback: Users can now provide feedback to agent LLM calls within the Agentic table. Right click on your LLM populated cells and select Feedback from the dropdown.

Admin Configure whether or not column edits trigger row re-run: Admins can now control whether editing a column triggers a re-run of the agent for a respective row’s dependent cells. Previously, this behavior was hard-coded. A new rerun_trigger parameter (default: true) enables or disables reprocessing of all rerunnable columns in that row when a column is edited. The default maintains backward compatibility, so existing Agentic Table configurations do not need updates.

• Backward-compatible by default: Existing Agentic Table setups require no changes.

• Expanded trigger capability: Columns like Answers, References, and Hallucination Score—which were previously excluded—can now trigger re-runs if edited.

• Non-editable columns do not re-run: Columns with the attribute editable: false, such as references and hallucination check, can not trigger a re-run.

For more details, see the documentation.

🔧 Improvements

Admin Sheet Review Status Visibility Is Now Configurable: The overall review/verification status now can be fully hidden when the Review Status column is removed from the sheet configuration. More details are available in the documentation.

Unique AI

🔧 Improvements

User Enhanced Token Management for Multi-Tool Agentic Loops - Token handling has been optimized to manage large content from multiple tool calls more efficiently, removing the slow iterative reductions that previously caused delays when token limits were exceeded. The update scales better for high-volume responses — especially with large-context models like Gemini — by reducing redundant token computations and improving overall performance.

Admin Claude Opus 4.5 is now available in the approved model dropdown for UniqueAI - It has been added following an extensive quality evaluation, demonstrating exceptionally clear, well-structured, and easy-to-read responses. While slightly slower and more variable than GPT 5, it performs best across both simple and complex prompts and occasionally adds relevant contextual details that enrich the overall answer quality.

🐞 Bug Fixes

Admin Logging for Orchestrator-Controlled Tools (e.g. Deep Research) - Tools that take over the orchestrator, such as Deep Research and the SWOT agent, are now correctly logged as tool calls in the debug information. This update ensures that they appear correctly in the analytics reports.

Admin Enhanced Tool Usage Logging - The side window with step logs now records both which tools were triggered and how many times each tool was invoked within a loop iteration (e.g. Internal Search (3x)). This update provides clearer insight into tool-usage patterns, and improves overall execution observability.

User Duplicate Step Logging in Deep Research Resolved - A bug that caused some Deep Research steps to be logged twice has been fixed.

User Fix for Chat Crashes Caused by Excessive Tool Calls - A bug that caused UniqueAI to crash when more than five tool calls were triggered by the orchestrator in one loop iteration has been resolved. The default limit is now set to 10 tool calls per loop iteration, and this value can be configured in the loop agent’s advanced settings.

Language Model

New Model Deployments:

Anthropic Claude Opus 4.5

Admin Claude Opus 4.5 is available through LiteLLM using the following enum:

• litellm:anthropic-claude-opus-4-5