Skip to content
Copyright © 2022-2026 basebox GmbH, all rights reserved.
Licensed to be used in conjunction with basebox, only.

// what's new

Versions

Applies to

Product: Cloud · Server · Audience: Everyone

Every basebox version since 1.1 is listed here, newest first. Each version starts with the essentials in a few sentences, followed by the complete list of changes to expand. What comes next is in the roadmap; which components belong to which version is in the version matrix.

1.9 · 20 October 2026: connectors per app, blocked terms and documents as files

Covers 1.9.0 to 1.9.3.

  • Connectors can be set per app, and every user sees their connectors on a page of their own.
  • Blocked terms hold back questions and answers that contain a forbidden word.
  • The assistant delivers Word, PowerPoint, Excel and PDF files, generated inside your installation.
  • Every citation shows how strongly the source supports the statement.
  • The admin dashboard is new, with estimated costs, limits and alerts.
  • Stored credentials are encrypted in the database, and the SBOM is under Version Info.

Before upgrading to 1.9

  • From 1.9.2 on, basebox starts only with a master key for stored credentials (AISRV_MASTER_KEY). The Docker Compose installer generates it; every other installation has to set it before upgrading. The key belongs in every backup, since without it every stored credential is lost.
  • 1.9 adds nineteen database migrations, applied automatically at startup on installations with automatic migration enabled. Existing connectors and organisation settings are carried over; the new switches default to the previous behaviour.
  • Locally run models use a newer inference engine. Graphics cards older than the Turing generation are no longer supported.
  • Users of the Nextcloud connector now sign in with their username and an app password and enter both once more.
  • The OpenAI-compatible API rejects a message with an unknown role with 400.
All changes in 1.9

Added

  • Your connectors page - Every user has a page listing the connectors available to them, showing which ones still need a sign-in and when each was last reached. Signing in happens in one dialog wherever a connector is met, and connectors are switched on and off right from the chat.
  • Connectors per app - App authors decide which connectors their app offers, on a new Connectors tab in the app editor, and the chat only shows those. Administrators choose per connector whether it is available in every app or only in apps that opted in.
  • Always-on connectors - A connector that needs no login and nothing from the user, such as the Document Generator, can be switched on for everyone by an administrator. It then simply works and no longer has to be found and enabled in the connector menu.
  • New connectors - CentralStation CRM, ISiK FHIR, roXtra, SharePoint and Zammad. The Atlassian connector is split into separate Jira and Confluence connectors.
  • Blocked words - Administrators keep a list of words and phrases that must not appear in a chat. Unlike instructions to the assistant, this is enforced: a question or answer containing one is withheld and the user is told that the content is not allowed under the company guidelines.
  • Your organisation's colour - Administrators set the primary colour used for buttons and highlights, with a warning when a shade would be hard to see in light or dark mode.
  • Document Generator - The assistant can turn what it writes into a Word, PowerPoint or Excel file, into the LibreOffice equivalents, or into a print-ready PDF. Files appear as a download link right next to their name in the answer, the link keeps working when you come back later, and nothing leaves your installation to produce them.
  • How well an answer is grounded - Every citation is labelled High, Medium or Low for how strongly the cited passage supports the statement, optionally with the exact percentage, and the chat says how many documents were searched and how many passages matched. Administrators tune the cut-offs.
  • A new admin dashboard - One filter bar (range, org unit, app, channel, model) applies to every panel, with CSV export: key numbers with their trend, a timeline with a marker for each new version, usage by app and by model, and how usage is spread across users, counted and never naming anyone. On top come an estimated cost from a per-model price table, and limits per organisation that warn under "Needs attention" but never block anything, plus flags for oversized requests, unusually large conversations, jumps after an upgrade and unusual access.
  • Copy button on code blocks - Hovering a code block in an answer shows a button in its top right corner that copies exactly its text. In a block taller than the screen it stays in the visible part; on touch screens it is always shown.
  • Redesigned emails - The invitation, renewed invitation, welcome and organisation confirmation mails share one design in the basebox colours, and their subjects say what they are about.
  • SBOM - The About page links the software bill of materials (SBOM) of every component and offers it as a machine-readable download, in CycloneDX 1.6 following BSI TR-03183-2. Every shipped container image has its own SBOM, and the on-premise installer package carries an SBOM of the whole installation, usable without network access.
  • Sources from web pages are linked - When a cited passage comes from a scraped web page, the citation leads to that page.
  • Beta features - Administrators unlock work-in-progress features and connectors for their organisation on the Organisation page; they are marked "Experimental".
  • Second factor by text message - Where two-factor authentication is switched on, users can receive their code by SMS as well as through an authenticator app. Two-factor authentication can also be required for individual groups only.
  • Chat history with search - Past conversations have their own page, can be searched and filtered by date, and can be deleted all at once.
  • Paste images into the chat - A screenshot or any other image on the clipboard can be pasted straight into the chat as an attachment.
  • Retention controls - Uploaded documents can be deleted automatically after a chosen number of days, and audit log retention changes take effect immediately without a restart.
  • Web search where you want it - Administrators decide whether web search appears as a toggle in every chat or as a dedicated Web Search app in the sidebar.
  • Clearer connector errors - When a connector call fails, the message says what happened and what to do, with technical details and a reference ID one click away for support.
  • A notice when no answer arrives - If the assistant stops without answering, the chat says so and offers to try again with less thinking.
  • The document list shows who uploaded each file and when, and the reasoning of a model streams live alongside connector activity.
  • More models – Each organisation's administrator picks which of the offered models the organisation answers with, in the Default model section of the Organisation page; an app can still use its own model. Claude Opus 5.5, Sonnet 5.5 and Haiku 5.5 are supported with thinking. Several providers can be offered side by side, including EU-hosted ones, and each model shows the region it runs in. Image generation gains a second route through Black Forest Labs' API with an EU endpoint.

Changed

  • Connector administration redesigned - Connectors are shown as cards with a single settings panel, the settings are asked as a few plain questions, adding a connector takes fewer steps, and each connector lists what it can do. The page has two tabs, Yours and Discover; a connector is switched on only once its settings are saved, and straight after adding one only its sign-in section opens. Connector names and descriptions are translated.
  • Server addresses are a setting - The email connector's mail server and the Nextcloud address are entered per organisation in the connector settings, so one connector serves organisations on different servers.
  • Connectors stay current after an upgrade - At every start basebox re-syncs the connectors with their servers: new, renamed or dropped tools and changed connections arrive without an administrator saving anything.
  • No organisation signup on-premise - An on-premise installation has exactly one organisation, so it no longer has a signup page.
  • Stricter OpenAI-compatible API - A message with an unknown role or without content is rejected with 400 instead of silently being treated as a user message; developer is accepted as the system role.
  • Settings belong to each organisation - Connectors, the support address, the welcome message and the groundedness cut-offs are kept per organisation, so on an installation shared by several organisations each administrator manages their own. Existing values are carried over on upgrade.
  • Newer inference engine - Local models run on a newer version of vLLM that supports more recent model architectures. Graphics cards older than the Turing generation are no longer supported.
  • Everything that used to be called "MCP" in the interface is now a "connector".
  • Finished reasoning collapses to a short "Thought for X steps" line that can be expanded.
  • The translator keeps its history automatically, remembers the target language, offers more languages and has a stop button.
  • An organisation is no longer limited to 100 custom apps.
  • Logging out asks for confirmation.
  • Many small refinements: plainer wording in the model selector, larger targets in the connector menu, a search icon for the chat history, traffic-light colours for document status, and a clearer distinction between documents and passages in the sources line.

Fixed

  • The usage overview counts everything that was used: speech-to-text, generated pictures, chats that call connectors, and answers cut off at the length limit. Chats with vLLM and other OpenAI-compatible models are counted again; usage from before the upgrade cannot be recovered.
  • Saving your own credentials, or an administrator switching a connector on, no longer switches it on for users; each user switches it on themselves.
  • Web search stays on in the window where it was switched on and is no longer switched off by another window. A search call no longer appears as raw text in the answer; it is run.
  • The chat's "+" menu makes it visible that it holds more than five connectors, and the list of connectors used sits at the end of the answer instead of covering part of it.
  • The Nextcloud connector works again; users sign in with their username and an app password and have to enter them once more. The Atlassian connector's sign-in test works.
  • Connectors whose server moved during an upgrade, such as the calculator, work again without being added anew.
  • Connectors each user signs in to are no longer flagged as "Needs attention".
  • Editing a sent prompt no longer makes the chat jump, and the drop area shown while dragging a file reliably disappears once the drag ends.
  • App descriptions on the app card are shown in full lines.
  • Saving a connector's settings no longer clears its title and description.
  • Connectors that need nothing from the user no longer ask for credentials, and password managers no longer fill in a connector's token field.
  • Cited passages containing tables, images or links render as text instead of raw markup.
  • The chat scrolls all the way to the end of a finished answer, and the page-size selector in the file list works as expected.
  • The sidebar stays open while an app is being edited.
  • Generated PowerPoint files open without repair prompts.

Security

  • The TLS library used by all backend services was updated to close RUSTSEC-2026-0285, in which TLS 1.3 handshake messages at the wrong encryption level were accepted. Connections could not be altered or completed by an attacker, but the specification violation is fixed. A withdrawn cryptography crate version was replaced at the same time.
  • Stored credentials are encrypted - Users' connector credentials, a connector's shared account and model provider keys are stored encrypted in the database. Credentials saved earlier are encrypted in the background at startup. aisrv starts only with the key AISRV_MASTER_KEY: the Docker Compose installer generates it, every other installation has to set it before upgrading. The key belongs in every backup, since without it every stored credential is lost.
  • The TLS library was updated again: under RUSTSEC-2026-0349 an unauthenticated server reply could make later outbound connections to the same host fall back to classical key exchange without post-quantum protection; connections stayed authenticated and encrypted. Three further advisories for the same library do not affect basebox.
  • The zip library was updated (RUSTSEC-2026-0367 to 0369). The flaws concern extracting archives to disk, which basebox does not do; no version was exploitable.
  • Third-party actions used in the build pipeline are pinned to fixed versions, so a re-pointed tag cannot change what gets built.

1.8 · 19 August 2026: connectors, web search and a new design

  • Connectors link basebox to wikis, mailboxes and ticket systems; everyone signs in with their own account.
  • Web search is optional and asks for consent before the first search.
  • Users choose how thoroughly the assistant thinks before answering.
  • Answers based on your documents name those documents.
  • The audit log has its own area with filters and CSV export.
  • The interface has a new design, and documents are processed faster.
All changes in 1.8

Added

  • Connectors - basebox can now be connected to the systems your team already works with, such as wikis, mailboxes and ticket systems. Everyone signs in with their own account, so people only ever see what they are allowed to see. Connectors are enabled per app, can be tested with a single click, and show whether the connection is currently healthy. The assistant knows which connectors are available and uses them when they help to answer a question.
  • Custom welcome message - Administrators can tailor the welcome title and message shown to their organisation.
  • Web search - An optional web search lets the assistant look things up online. It is switched off by default, asks for the user's agreement before the first search, and can be enabled or disabled for an entire installation.
  • Thinking modes - With models that support it, users can decide how thoroughly the assistant should think before it answers, and follow its train of thought in a panel that can be collapsed at any time. Administrators set the default for individual apps and for the whole installation. Thinking is also available through the OpenAI-compatible API.
  • Sources and citations - Answers that draw on your own documents now name those documents, and a click on a source opens the document itself. The chat also shows how much of the knowledge base was searched and how well each source matched the question.
  • Company-wide and personal instructions - Administrators can define instructions that apply to everyone in the organisation, for example a house style or a required language. In addition, every user can save personal instructions for their own conversations.
  • Audit viewer - A new administration screen shows the audit log with paging, filtering and export to CSV. Both the retention period and the level of detail are configurable: the log can record questions only, or the assistant's answers as well. Whenever conversation content is being recorded, everyone sees a clearly visible notice in the chat.
  • Limits per organisation - Upload and usage limits can now be set separately for each organisation.
  • More file types, better text recognition - Word .doc files, e-mail messages, SVG and EMF graphics are now supported. Scans, photographs and Word files that contain nothing but images are read using text recognition, which uses a graphics card where one is available and falls back to the processor automatically. Entire folders can be added to a knowledge base by drag and drop or from the file dialogue.
  • Translation history - The translator keeps your earlier translations so you can return to them.
  • Longer dictation - Voice input accepts recordings of up to one hour, and audio attachments show their transcript.
  • Everyday conveniences - Conversations can be renamed, the sidebar can be collapsed to give the chat more room, apps can be filtered and searched by author, administrators can edit and delete the supplied default apps, users can delete their own account, and the mail configuration can be tested before it is saved.
  • Usage overview - The dashboard reports usage of the OpenAI-compatible API separately from the web interface, and the About page shows the active model, how much context it can hold, and whether speech recognition and the API are available.

Changed

  • New design - A new look and feel across the whole application, with reworked organisation and profile pages, one Save button per page that warns before unsaved changes are lost, clearer switches in place of drop-down lists, and calmer progress messages while an answer is being written.
  • Faster document processing - The pipeline that reads your documents has been rebuilt from the ground up and is noticeably faster and more accurate, with a new text recognition engine.
  • Every document is searched - A search now considers all files in a knowledge base instead of only the first ten.
  • App store retired - Apps are managed directly in the app list.
  • The OpenAI-compatible API is always available - It no longer has to be switched on; it is part of every installation.
  • Web search starts switched off in every new conversation, so using it is always a deliberate choice.
  • Personal instructions accept up to 1000 characters and show examples of what to put there.
  • A connector's own website opens in a new window.
  • While an answer is being written, the chat shows which connector and which tool are in use, and how long it has been running.

Fixed

  • Uploads - Files with the same name are replaced automatically, a deleted file can be uploaded again straight away, the original file name is kept when uploading folders, files that are too large produce a clear message naming the real limit, and sending is blocked while uploads are still running.
  • Attachments - The preview no longer covers the input area when many files are attached, and files are accepted up to the allowed number with an accurate count of those left out.
  • Chat - Output catches up immediately when you return to a tab you had left, long answers are no longer cut short, the input area and the "AI can make mistakes" note stay visible when announcements are displayed, and messages containing emoji no longer cause errors.
  • Prompt input - Unsent drafts are restored when returning to a conversation, and Markdown lists continue automatically while writing.
  • Documents - Images and image-only PDFs are processed reliably, citations work in every interface language, and a source link opens the document instead of starting a new conversation.
  • Speech - The language of a recording is now detected automatically instead of following the profile setting, recordings that are not WAV files are transcribed correctly, and recordings that are too large are rejected with a clear message.
  • API - An invalid key returns a precise authentication error, no artificial instructions are added for Anthropic models, and there are clear messages when a model is unknown, a request takes too long, or a question exceeds what the model can hold.
  • On-premise operation - User synchronisation with the identity provider is more robust, audit events include the client address, failing token requests are logged with the address concerned, database updates can be repeated safely, and the identity provider and web server configuration used when upgrading from 1.6 or 1.7 has been corrected.

Security

  • Connector credentials are stored write-only: they are never sent back to the browser, are hidden in log files, and are never filled in automatically. A badge shows when credentials have been set.
  • Supply-chain scanning was added to the build pipeline, and dependencies were updated to address published security advisories.

1.7 · 25 April 2026: OpenAI-compatible API and group management

  • Tools built for the OpenAI API work with basebox: the OpenAI-compatible API.
  • The Swagger UI at /v1/api/docs tests the API in the browser.
  • With the message of the day, administrators reach all users.
  • Group management has live search and clearer permissions.
  • New apps of your own are not shared at first.
All changes in 1.7

Added

  • OpenAI Compatible API - Full OpenAI API compatibility for direct integration with existing tools and applications
  • Group Management - Backend support for user group management and organization
  • Swagger UI Documentation - Interactive API documentation with testing capabilities (optional feature)
  • Tool Support - Enhanced tool calling capabilities with auto-handling for vllm endpoints
  • Message of the Day - Admin configurable system-wide announcements with frontend and backend support
  • Server Readiness Check Script - New operational tooling for checking server health and readiness
  • Pagination Controls - Added 'Items per Page' control to pagination component for better data browsing
  • Integration Test Infrastructure - Comprehensive integration test framework for backend services
  • OpenClaw Integration - Added integration guide and support for OpenClaw
  • Enhanced Audit Logging - Comprehensive audit system with SSE streaming, configurable logging, and partition management

Changed

  • Major Architecture Update - Refactored chat completions and models for better maintainability and single responsibility
  • App sharing: new custom apps are initially not shared; you have to explicitly share them if you want others to be able to use them
  • Improved streaming support and message role handling (tool, function, assistant)
  • Enhanced Vertex AI mode with full tool call support
  • Updated integration tests for both streaming and non-streaming LLM interactions
  • Improved on-premise mode configuration (no longer requires app_id for LLM config)
  • Enhanced price per token calculations and model handling
  • Updated authentication and license handling in integration tests
  • Group Management UI Improvements: Enhanced group creation UX with keyboard shortcuts, better error handling, and improved member management
  • Permission Management: Streamlined permission changes with direct dropdown access and clearer terminology
  • Search Experience: Implemented live search with debounce for user and group lists
  • Mobile UI Enhancements: Improved action button visibility on touch devices and better mobile app editor layout
  • Database Migrations: Enhanced migration system with single transaction support and rollback on failure
  • Audit Infrastructure: Major improvements to audit logging with blocking behavior, better error handling, and automatic partition creation
  • Model Endpoint Exposure: Enhanced vLLM endpoint integration to expose all available models
  • Session Management: Improved Keycloak session handling and logout API behavior
  • Keycloak Upgrade: Updated Keycloak to version 26.5.6 for improved security and features

Fixed

  • Fixed Anthropic API not properly handling "tool" type calls
  • Improved model validation for cloud version deployments
  • Fixed test database configuration and password management
  • Enhanced error handling and logging throughout the system
  • Fixed nginx routing configuration
  • Added /v1 location block to expose /v1/api/docs (Swagger UI), /v1/models, and /v1/chat/completions endpoints
  • Fixed missing /rest location block in TLS nginx configuration
  • Updated nginx Dockerfile to include basebox-application-tls configuration file
  • LLM Response Parsing - Fixed response parsing failures and improved error transparency
  • Database Migration Error Handling - Enhanced migration error reporting with detailed error chains
  • Context Length Errors - Better error messages when prompts exceed model context limits
  • Socket Security CLI - Fixed CLI arguments and Gitea API compatibility issues
  • On-premise OpenAI Configuration - Fixed to use organization-level subdomain instead of host
  • Keycloak User Synchronization - Fixed user synchronization when user already exists in Keycloak
  • User Field Management - Fixed read-only user fields in Keycloak integration
  • Mobile UI Issues - Fixed action buttons not showing on mobile/touch devices and improved chat input scroll behavior
  • STT Configuration - Fixed STT UI display respecting enableSTT configuration variable
  • File Status Management - Improved file status handling and replacement workflows
  • App Menu Options - Added dropdown with app edit and app sharing options for better navigation
  • Audit Export - Fixed audit query and export API functionality

Security

  • Security Scanning - Added Trivy frontend security scanning for vulnerability detection
  • Resolved security workflow compatibility issues

1.6 · 7 March 2026: app sharing, speech input and reasoning models

  • You decide who may use or edit your own apps.
  • basebox understands speech and audio files in the chat.
  • Reasoning models are supported.
  • Invitations are valid for 14 days instead of 7.
All changes in 1.6

Added

  • App sharing: specify who can use or edit your custom apps (feedback wanted!)
  • Load more button for chat history in sidebar
  • Show warning if users are being invited without SMTP configured (on-premise)
  • Delete chat from history only after confirmation
  • Reasoning model support
  • Added backend and frontend SBOM links to credits (see main menu -> About -> List of Open Source software used in basebox)
  • Speech to text (STT) - talk to basebox!
  • RAG file list in app editor now supports pagination and searching by name
  • Audio file extraction: you can now drop audio files on the chat
  • Runtime frontend configuration for on-premise installations

Changed

  • Change layout of chat interface (prompt box)
  • Changed inference engine to latest vllm (TGI is in maintenance mode)
  • Invitations are now valid for 14 days (used to be 7)

Fixed

  • Fixed tooltips not reloading after changing the language setting in the user profile
  • Layout improvements (spacings etc.)
  • Fixed wrong login theme for on-premise installations
  • Fixed IndexedDB versioning error
  • Fixed various UI bugs (buttons not being disabled during load etc.)
  • Display on-premise config errors with a nice error message
  • Fixed RAG file list pagination issues
  • Fixed chat state after stopping assistant's answer generation
  • Various fixes and improvements and fixes for using basebox on a mobile device
  • Improved scroll behaviour when editing a chat message
  • Improved security when displaying AI-generated content
  • Fixed language in app list not being updated after user language was changed
  • Fixed various UI bugs and peculiarities
  • Fixed organisation logo editing and saving issues

Security

  • Security related fixes (dependency updates)

1.5 · 12 December 2025: licence management, split view and this documentation

  • Server installations have licence management in the admin area.
  • Two apps can be used side by side in a split view.
  • The basebox Documentation is online.
All changes in 1.5

Added

  • License management page in admin area for on-premise installations
  • Technical documentation site now online, https://docs.basebox.ai
  • Split screen UI

Changed

  • Improved error reporting in chat
  • Removed first and last name fields being required so language can be set without entering a name
  • Allow more file types for chat and app uploads
  • Performance improvements
  • Added static file server to aisrv for on-premise installations
  • Reverted app sharing scope to the entire organisation (as in pre 1.0.0 versions); an update with user and/or group scope sharing is currently being tested.

Fixed

  • App editor showed no more than 10 RAG files, even if the app had more
  • Improved error reporting for context-too-large errors
  • Fixed copy/paste for rich text content
  • Improved document extraction retry logic
  • Fixed file counter and limit handling when uploading RAG files
  • Fixed app RAG file deletion issue
  • Fixed display status updates during RAG file processing
  • OCR stability improvements
  • Fixed text entering in split screen apps
  • Fixed various minor UI issues
  • Fixed potential database initialization issue for on-premise (storesrv)
  • Various minor layout/UI fixes and improvements
  • Temperature control in app editor fixed
  • Improved error reporting
  • Fixed case sensitive file type filter in chat (accepted .pdf but not .PDF)

Security

  • Various security related fixes

1.1 · 28 October 2025: the first 1.x release with a new interface

  • The interface is redesigned.
  • Images and image-only PDFs are read with text recognition.
  • A new pipeline processes documents faster and more accurately.
All changes in 1.1

Added

  • OCR support - drop images and image PDFs into the chat

Changed

  • New layout, look & feel
  • Ported the entire frontend from React to Vue.js, all TypeScript
  • New document extraction pipeline with vast improvements (accuracy and speed)
  • Improved app editor
  • A thousand other details

Fixed

  • Many bugs and issues