> ## Documentation Index
> Fetch the complete documentation index at: https://docs.varios-ai.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Changelog

## 2026.06.10 ES

*date: 2026-07-17*

### Bugfixes

* Model sensitivity label selections persist correctly across save and reload
* Filtering knowledge base documents by embedding status no longer fails when no documents match

## 2026.06.9 ES

*date: 2026-07-15*

### Bugfixes

* Action buttons on assistant tiles are always visible (not hover-only), making them reliably accessible on all screen sizes
* Newly created assistants no longer disappear briefly from the list after saving
* Prevents chat failures after knowledge base retrieval when chunk metadata has no filename and the linked document resource is missing

## 2026.06.8 ES

*date: 2026-07-14*

### Bugfixes

* Demo remaining credit is refreshed again after chat usage
* Effective DLP state is shown correctly on the assistant data protection tile when the model enforces DLP
* PostgreSQL/pgvector image is now built for the x86-64-v3 baseline (AVX2) instead of the native CI CPU, preventing crashes when images built on one x86 host run on another
* `update.sh` now warns before installation when the CPU lacks AVX2, as this can cause pgvector-related PostgreSQL crashes

## 2026.06.7 ES

*date: 2026-07-07*

### Bugfixes

* Support for arbitrary gpt-image-2 resolutions as image size
* Human-in-the-Loop approval prompt no longer reappears when reloading the chat during a running connector execution
* Improved readability of the Human-in-the-Loop info box in Light mode
* Compliance dashboard is accessible with the ComplianceAdmin role without requiring additional admin rights
* Connector media (images and videos) are no longer shown twice when already embedded in the chat content
* WebSocket connection no longer drops unnecessarily during long streamed answers
* Fixed embedding provider service-state checks (including IONOS); error messages are displayed more reliably

## 2026.06.6 ES

*date: 2026-07-03*

### Bugfixes

* Human-in-the-Loop works again for code connectors whose identifier starts with "f" (Responses API)
* Connector costs in the dashboard are attributed to the correct connector again (instead of "deleted")
* The loading animation appears again after confirming a Human-in-the-Loop prompt while the connector runs

## 2026.06.5 ES

*date: 2026-07-01*

### Bugfixes

* Parallel chat sessions work again: you can start or continue a new chat while another chat is still sending a message
* Overlapping DLP analyzer spans no longer produce nested type markers in the DLP preview

## 2026.06.4 ES

*date: 2026-06-30*

### Bugfixes

* Human-in-the-Loop now also works for OpenAPI and MCP connectors
* Reasoning and assistant code blocks are removed from AI-generated system prompts
* Suggested assistants remain invisible to users, even if they are assigned to a group
* Faulty Mermaid diagrams no longer display internal error blocks

## 2026.06.3 ES

*Datum: 2026-06-25*

### Bugfixes

* Corrects the validation of the data protection overlay

## 2026.06.2 ES

*date: 2026-06-25*

### Bugfixes

* Fixed issue with too many files in knowledge bases
* Chat messages containing only attachments and no text can now be displayed and edited.
* Thought processes without a title are now displayed correctly again.
* Fixed German translations.
* The custom assistants heading is displayed more reliably.

## 2026.06.1 ES

*date: 2026-06-18*

### Improvements

* Adjusted WebSocket logs.
* Transcribe errors are now logged to the System Logs.

### Bugfixes

* The Human-in-the-Loop checkbox is now displayed.

## 2026.06 ES / 2026.6.0 ES

*Date: 2026-06-17*

* The release and versioning model has been further developed: Extended Support Release and Monthly Release

### Features

* Human-in-the-loop for connectors
* AI-assisted editing of the system prompt when creating and editing assistants
* New dashboard for service statuses of connected AI services
* Charts and diagrams can now be rendered directly in chat using Mermaid
* VARIOS AI n8n-Node

### Improvements

* Cleaned up and better organized privacy settings
* Group-level budget and usage in the group view
* Made the AI model used to generate embeddings configurable in knowledge bases
* Enabled configuring cost thresholds by context length in the model settings
* Added helper text below the chat input
* Made the start heading in chats configurable

### Bugfixes

* Various minor display and text issues have been fixed (including labels and translations).
* The chat list now updates correctly when creating a new chat, without requiring a full reload.
* Stability during login/identity verification has been improved.

## 1.164.27

*Date: 2026-06-09*

### Improvements

* Output token limit can now be configured in the model settings

### Bugfixes

* Fixed resource cleanup when deleting chats
* Malformed XML files will now be treated as plain text

## 1.164.24

*Date: 2026-05-20*

### Improvements

* Video connectors for Google Veo and xAI Video added
* WebSocket connection and recovery have been made more robust
* Reasoning and code blocks removed from speech output and usage of fake data adjusted

### Bugfixes

* Empty object values are now removed from the tool definition for MCP tools
* `lastImagesCount` in `GoogleImage` is now correctly clamped to the range 0–3
* DLP option for the embedding model is now correctly reloaded in the knowledge bases
* Conversion of math delimiters within regular expressions no longer takes place
* File sensitivity label checks are now performed case-insensitively
* Source display for knowledge bases is now also available for standard users

## 1.164.11

*Date: 2026-04-22*

### Improvements

* Introduces a new Entra app role to manage login to VARIOS AI

### Bugfixes

* Fixes responses for reasoning models

## 1.164.8

*Date: 2026-04-15*

### Improvements

* Expands the API with the Open Responses format
* Enables deleting used models
* Enables hiding Code Interpreter output from the PDF chat export
* Introduces a blacklist for group users

### Bugfixes

* Fixes issues with MCP tool synchronization
* Fixes errors when saving assistant cards
* Fixes errors when overriding reasoning effort

## 1.164.7

*Date: 2026-04-08*

### Improvements

* Restructures the admin settings
* Allows disabling base prompts in the prompt library
* Optimizes page navigation (jump to a specific page) in the PDF viewer
* Adds a new configuration option for `maxCompletionTokens`

### Bugfixes

* Prompt Library: Tags and categories are now correctly filtered by the selected language
* Fixes an issue that caused incorrect overwriting of model parameters
* Resolves a layout issue in the fullscreen modal of the PDF viewer

## 1.164.2

### Features

* Source display when searching knowledge databases
* Overriding model parameters in the wizards
* Allow-list improvement: Free-form field for bulk uploading of regular expressions
* DLP enhancement: Detection of VARIOS AI users (first name, surname)
* Extension of the VARIOS AI /chat/completions API to include tools, among other things
* Prompt library

### Bugfixes

* API errors with some embedding providers corrected
* System messages removed from API (debug flag in settings)
* List of costs and DLP interventions per API key
* Confidentiality labels can be maintained again
* Chats containing used files can be deleted again
* The LogViewer is scrollable again

## 1.156.46

### Bugfixes

* Connectors can be re-created
* Fixes errors when saving settings
* MCP configuration fix
* Scales images of wizards, models and connectors down
* No error message for OpenAPI YAML configurations

## 1.156.41

### Bugfixes

* Removes incorrect display that files are being processed
* Costs per API key are displayed in the dashboard
* PDF export does not support non-Latin characters
* Reasoning and code analysis can be deactivated in PDF export
* Minor bug fixes

## 1.156.38

### Bugfixes

* Removes incorrect logging
* Minor UI corrections
* Expired links in old chats are regenerated
* Corrects incorrect URL handling in embedding models
* Removes internal system messages in the API

## 1.156.34

### Bugfixes

* Slash at the end of endpoint URLs is now optional
* Image is copied when copying

## 1.156.33

### Features

* Introduces voice - text-to-speech & speech-to-text
* Makes hybrid RAG adjustable
* Enables chat switching without losing the entered message
* Makes titles, custom logos, and favicons editable
* Enables different sources for the knowledge base with automatic synchronization
* Adds action logger for connectors
* Allows chats to be downloaded as PDFs
* Adds connector costs to the dashboard
* Allows multiple chats at the same time
* Introduces user soft deletion and restoration
* Improves file upload logic and prompt configuration
* Makes chat list width customizable
* Improves Keycloak user synchronization
* Models and wizards can now be duplicated
* Adds an expert view for advanced settings
* Updates initial instance data
* IONOS Flux is supported
* Adds /models endpoint to API
* Enables Redis as cache backend
* Enables logging via Syslog
* Removes deprecated OpenAI Assistant API

### Bugfixes

* Fixed deletion of users when a group is deleted via SCIM
* Fixed code block size in mobile view
* Optimized chat loading process
* Optimized LLM error messages
* Now correctly saves OpenAPI baseUrl and endpoints
* Files can now only be removed using the “X”
* Standardized file upload validation
* Removed model backdrop when clicking next to it
* Deletes users via SCIM when a group is deleted
* Added command to delete orphaned users
* Improved loading indicator
* Fixed display for incorrectly configured models
* Optimized OpenAI vector store search
* Corrects handling of DLP-encrypted chat files
* Improves MCP implementation
* Corrects translations of DLP entities
* Fixes truncated text in model lists
* Enables larger input field
* Enables code interpreter without files
* Avoids unnecessary re-rendering
* Fixes various translation errors
* Summarizes costs for deleted entities in the dashboard
* Chat messages may contain a maximum of 50,000 characters
* Maximum number of results can be defined in the Microsoft 365 connector
* Uses Markdown instead of HTML in the Microsoft 365 connector
* Prevents downgrade to Stable if current version > Stable version
* Fixed PDF view (Think blocks and documents)
* Fixes minor bugs

## 1.140.45

### Bugfixes

* Fixes bugs when processing SVGs
* Adds missing pagination in the dashboard
* Minor bug fixes

## 1.140.41

### Bugfixes

* Corrects deletion of users when a group is removed via SCIM
* The browser language is taken into account during the first login
* Improves Keycloak user synchronisation
* OneDrive (Microsoft365) files are passed to the LLM as Markdown text

## 1.140.33

### Bugfixes

* Corrects cost tracking
* Fixes display errors for selected OpenAI/MCP endpoints
* Replaces DLP entities in the title and summary
* Introduces a lock during user synchronisation, which prevents the process from running multiple times

## 1.140.31

### Bugfixes

* Fixes errors through encrypted reasoning by OpenAI

## 1.140.30

### Bugfixes

* Optimised MCP request
* Enhanced Google Nano Banana connector
* Corrected Gemini token count
* Minor bug fixes

## 1.140.24

### Bugfixes

* Data protection text consent can be obtained again from all users
* Option to upload your own logo
* Represents Gemini's Think Block
* Corrects translations
* Fixes style errors in the data protection text
* Optimises vector search with the OpenAI API
* Minor optimisations

## 1.140.11

### Bugfixes

* Corrected copy & paste from Office products (text instead of image)
* Minor bug fixes
* No longer sends the user hash for non-streaming

## 1.140.9

### Bugfixes

* Fixes errors when querying assistant files
* Does not transfer documents (only images) to OpenAI when local knowledge base is active
* Corrects minor error in the OpenAI Responses API

## 1.140.8

### Bugfixes

* Corrects MimeType for encrypted document text
* Optimises knowledge base search

## 1.140.7

### Features

* Introduces DLP encryption for files
* Enables use of Google Gemini 3 Pro and Gemini 3 Pro Image (Nano Banana)
* Creates loggers for connectors

### Bugfixes

* Corrects chat history in mobile navigation
* Corrects minor OpenAPI/MCP errors
* Corrects link to assistant card
* Closes the system update modal correctly
* Corrects the calculation of user costs
* Extends the time for the vector database

## 1.139.10

### Bugfixes

* Increases timeout for OpenAI vector search
* Also uses default SSL certificates in addition to the HTTP proxy certificate
* Corrects the calculation of user costs
* Corrects bugs in OpenAPI and MCP

## 1.139.9

### Bugfixes

* Fix bugs in OpenAPI / New ApiKey query option

## 1.139.8

### Bugfixes

* Corrects errors with “undefined” language
* Optimizes error messages
* Prevents double counting of DLP incidents

## 1.139.5

### Bugfixes

* Fixes the re-connect functionality for long-running requests

## 1.139.4

### Bugfixes

* Fix dlp tracking of files
* Optimizes reconnect of mqtt connections

## 1.139.3

### Bugfixes

* Fixes incorrect line breaks in code blocks

## 1.139.2

### Bugfixes

* Prevents establishing a new connection with every push
* Corrects the calculations of costInput and costOutput in the useModel hook for higher accuracy

## 1.139.1

### Bugfixes

* Improves the styling of the privacy editor
* Provides the missing “Admin” permission for models
* Optimizes the MarkItDown integration

## 1.139.0

### Features

* Introduces OpenAPI support
* Integrates MCP support
* Enables local document processing with the configured embedding model
* Displays costs per assistant, chat model, and embedding model in the dashboard
* Provides optimized logs (including separate audit, auth, DLP, SCIM, system, and security logs)
* Adds a compact compliance dashboard
* Implements hybrid RAG (combination of vector and full-text search)

### Bugfixes

* Enables the edit button in the assistants list for admins
* Adds DLP tracking in additional places
* Fixes the avatar display in the dashboard
* Hides the confirm button in the FileDLP modal when processing is not yet complete
* Improves error handling
* Uses encrypted reasoning in more requests
* Optimizes the ICAP DLP connection
* Adds new background colors for additional entity types
* Allows additional protocols in links
* Updates packages
* Fixes minor issues

## 1.134.38

### Bugfixes

* Provides the missing “Admin” permission for models

## 1.134.37

### Bugfixes

* Corrects the re-connect functionality for long-running requests
* Fixes minor bugs

## 1.134.36

### Bugfixes

* Fixes issues when deleting documents in knowledge bases
* Corrects HTTP proxy exceptions and certificate handling
* Updates MarkItDown to fix issues
* Adjusts translations

## 1.119.2

### Features

* Integrates a new OpenAI GPT Image model via connector
* Introduces a new, faster response API
* Provides new MySQL and Postgres connectors
* Adds new user and group pages
* Allows configuration of file types per model

### Bugfixes

* Fixes issues with login and scroll behavior
* Updates and corrects translations

## 1.114.13

### Features

* Automatically skips empty GDPR modal dialogs
* Introduces cache optimizations to improve performance

### Bugfixes

* Closes security-relevant vulnerabilities
* Removes the changelog entry regarding cleanup

## 1.113.13

### Features

* Introduces a new assistant view
* Provides a new MSSQL adapter
* Makes the application name variable in texts
* Allows resizing the system prompt (resizable)

### Bugfixes

* Optimizes the prompt cache
* Adds extra protection for token refresh
* Improves the theme build
* Fixes the UTF-8 issue in SearXNG and Bing
* Corrects chat width
* Secures SCIM endpoints
* Optimizes the stop function

## 1.112.18

### Bugfixes

* Closes security-relevant vulnerabilities

## 1.112.12

### Bugfixes

* Fixes a security vulnerability
* Corrects the UTF-8 issue in SearXNG and Bing again

## 1.112.7

### Features

* Enables knowledge bases
* Optimizes load times to improve performance
* Improves file handling
* Makes the assistants list automatically respond to available widths
* Introduces additional security prompts
* Improves Keycloak integration
* Stores the theme in the license key

### Bugfixes

* Cleans up code to improve stability
* Fixes various issues

## 1.111.11

### Features

* Re-enables creating custom assistants

### Bugfixes

* Fixes a rare issue with load order

## 1.110.5

### Features

* Adds a new Databricks connector
* Adjusts the ordering of connectors
* Optimizes Keycloak integration
* Updates initial data during installation

## 1.110.2

### Features

* Adjusts initial data during installation

### Bugfixes

* Fixes issues in DLP functionality

## 1.110.1

### Features

* Automatically replaces placeholders (currentDate/currentDateTime) in the system prompt

## 1.110.0

### Features

* Introduces an instance system prompt (applies always / before every assistant system prompt)
* Provides a best-practice system prompt template when creating new assistants
* Allows enabling connectors per group (custom assistants can use connectors)

### Bugfixes

* Performs minor optimizations and bug fixes

## 1.109.16

### Features

* Optimizes the loading animation
* Adds a MarkItDown URL setting
* Improves log configuration
* Fixes errors in demo data
* Resolves a performance issue
* Makes the OpenAI OrgKey optional
* Adjusts translations

## 1.109.8

### Bugfixes

* Fixes DLP errors when DLP enforcement is active
* Corrects issues on the home page
* Optimizes container settings

## 1.109.6

### Bugfixes

* Corrects color schemes
* Fixes an error with empty result pages in SeaXNG search

## 1.109.4

### Features

* Optimizes spacing in the DLP display
* Introduces a confirmation before sending the original text
* Adds the account ID to the license

## 1.109.3

### Features

* Adds missing translations
* Optimizes the handling of regular expressions in the deny list

## 1.109.2

### Bugfixes

* Fixes a DLP permission error

## 1.109.1

### Features

* Improves file handling

### Bugfixes

* Corrects DLP check errors
* Fixes minor bugs

## 1.109.0

### Features

* Highlights data recognized by DLP and allows the user to decide whether to proceed with or without DLP
* Provides an optional source reference as a badge in the text
* Introduces a license key check and sets the number of user and API licenses
* Adds a new top-p option to the AI model
* Shows favicons for search result sources
* Introduces a new “alpha” release channel for internal avency use
* Enables an (optional) regular Keycloak synchronization of users and groups
* Fixes the image lightbox in chat messages
* Updates various libraries

### Bugfixes

* Fixes minor issues

## 1.108.11

### Features

* Prepares license management

## 1.108.10

### Features

* Shows favicons for sources
* Improves Keycloak integration
* Increases the maximum number of files in the knowledge base
* Integrates Google Gemini, o3 mini and Azure DeepSeek R1 support
* Displays technical error messages
* Introduces an allowlist for DLP
* Enables editing own messages on mobile devices (swipe to the side)
* Optimizes OneDrive integration
* Allows removing the user identifier (chats will be deleted)

### Bugfixes

* Fixes minor issues

## 1.107.1

### Features

* Adds a think block in assistant messages
* Integrates DeepSeek R1 support

### Bugfixes

* Fixes minor issues