Changelog

​

Workspace

​

✨ Features

• GPT-OSS Model Now Available: You can now access our experimental GPT-OSS model in both 24GB and 48GB configurations, giving you more powerful AI capabilities for complex tasks. More information in experimental presets.

​

🐛 Bug Fixes

• Improved connection stability and reliability.
• Fixed a bug that could cause crashes when setting up new Agents Flows.
• Fixed a rare issue that could cause the AI to get stuck while processing requests.
• Fixed a bug where new agent flows couldn’t be created if a title wasn’t generated automatically.
• The system now chooses the correct language model for each task more consistently.
• Fixed an issue that affected data syncing between systems.
• Fixed an issue that occasionally affected image interpretation in PPTX files.

​

🛠️ Platform

• Use your own AI models (if needed): You can now plug in your own fine-tuned AI models without losing any of the platform’s features. This gives you more control and flexibility over how the AI behaves. . More information in customizing the default presets.
• Vision support improved: Fixed a bug that sometimes prevented AI tools from loading when visual (image) features were turned on.

​

Workspace

​

✨ Features

• Database Connector: Connect seamlessly to PostgreSQL, MySQL, and MSSQL databases and ask questions in chats, enabling intuitive data retrieval without writing SQL queries.
  • This feature is in Alpha phase and needs to be enabled.
  • Learn how to use it with the User Manual.
• Agent Flow - New menu options for an enhanced user experience:
  • Retry: Allows users to restart failed workflows with a single click.
  • Duplicate: User is taken a new workflow prefilled with the same data as the original

​

🐛 Bug Fixes

• When adding new members to projects search finds users by email.
• When the knowledge base had only processing files source selector showed 0 files instead of empty state.
• Omnisearch finds folders correctly.

​

🛠️ Platform

• API Refactor: Enhanced ZGPT API with automatic context management and direct tool integration, streamlining development workflows through intelligent request handling

​

🛠️ Platform

• Zylon becomes multimodal! Now our platform is able to manage several AI models running in parallel within the same instance.

​

Workspace

​

✨ Features

• Enhanced Slides Content Extraction thanks to run a Vision Language Model (VLM) in conjuntion with a Large Language Model (LLM). Please check the new alternative presets to understand how to configure and enable it. The brand new capabilities added are:
  • Intelligent Content Understanding: Implemented VLM-powered extraction that can analyze and understand complex visual content including slides, diagrams, charts, and images.
  • Contextual Analysis: VLM integration enables better comprehension of slide context, relationships between visual elements, and textual content within images.
• Display the last update time for Agent Flows.

​

🐛 Bug Fixes

• Support exporting DOCX files with custom elements (e.g. citations, knowledge sources).
• Resolved an issue where PDF exports had formatting problems with nested lists.
• Fixed an issue with partial answer parsing during streaming.
• Addressed an issue where the issue reporting panel displayed an empty date.
• Resolved a display issue in the artifact linking modal on Microsoft Edge.

​

Workspace

​

✨ Features

• 🔎 Omnisearch: Now able to search for content inside Agent Flows and Knowledge Base Documents
  • Follow the instructions to enable it
• 📝 Issue reporting: Users can now report chats to admin organizations when encountering issues with Zylon’s responses.
• 🎨 New Rendering Engine: Faster & Smoother Chats with Zylon
  • Zylon can now understand and render Mermaid diagrams directly in the chat
  • Enhanced readability of code blocks for better clarity and visual appeal
  • Chat responses now render smoothly without causing formatting jumps during streaming
• 💬 Enhanced default chat behaviour: Enforced use of semantic search when files are selected and no other tools are chosen
• 🔐 Audit logs control: Added the ability to turn off audit logs — follow the configuration guide to disable

​

🐛 Bug Fixes

• Ensure project is loaded before navigating to the first chat
• Prevent user to access removed chats by url and show warning when trying to access deleted resources (chat, projects)
• Fixed an issue with breadcrumbs in Omnisearch that caused accessibility issues
• Resolved an issue where numeric analysis was incorrectly disabled during document processing
• Fixed navigation to deleted projects by clearing the last visited project if it no longer exists
• Prevented users from navigating to deleted agent flows via Omnisearch
• Fixed broken code blocks in streaming when the programming language wasn’t detected

​

🛠️ Platform

• Ingest: A unified endpoint for ingestion that supports text, files, and URIs, including support for asynchronous ingestion.
• Reorganize the endpoint structure for better clarity by grouping related endpoints under a common base path.

​

✨ Features

• 📁 Folders: Organize your knowledge base effortlessly using folders for better management of directly uploaded documents
• 💬 Async chats: You can now close Zylon while your chat continues in the background — multitasking made easier!
• 🔎 Omnisearch: Instantly search through your content to quickly navigate to your resources: Documents, Chats, Agent Flows.

• 🔗 Links: Files linked in a knowledge base from another project now stay in sync with the original — no duplicates or outdated versions

​

🛠️ Platform

• 🚫 Deprecated endpoints in Low-Level API have been removed

​

Features

• Users can now create folder in knowledge base
• Users can now move folders and files inside knowledge base via Drag and Drop

​

Features

• New Agent Flow. Read & Extract.

​

Features

• Export Agent Flow to PDF/DOCX

• Resumability on chat (now the user is able to connect in realtime to a zylon response and the user can refresh or leave the app and comeback without stopping the answer)

​

Bugfixes & improvements

• Fixed a bug that prevented to leave project instructions empty
• Connections to database are properly closed to avoid reaching the limit
• Creating a new Claromentis integration was not possible due to an error

​

Features

• Tool Selector: The user has the ability to select Tools between available: Numeric Analysis, and Extended Thinking (only available with Qwen 3).
• Numerical analysis (currently in beta): Enables standard operations on tabular data in a textual format, including calculations such as averages, filtering, selection, and data visualization using plots. At present, this functionality is available only when a CSV file is selected from the knowledge base. For optimal performance, we recommend adhering to the following guidelines:
• A single header row should be included.
• Dates should be formatted as yyyy-mm-dd.
• Numbers should follow the format XXXXXXX(.XX).
• Empty cells should contain one of the following values: “nan”, “none”, “null”, "", “undefined”, “n/a”, “na”, “nat”.

​

Bugfixes & improvements

• Fixed a bug that prevented Qwen from generating plots in certain situations
• Fixed a problem with Assistant name when company context was set
• Better user experience in small devices in Analytics view

​

Features

• New model is available and set as default in all baseline presets. Take a look to Qwen 3 page to more detailed information.
• For small presets (24, 32 GB), our new default model is Qwen 3 14B.
• For bigger presets (48, 96 GB), our new default model is Qwen 3 32B.
• Extended Thinking: New thinking capabilities thanks to Qwen 3 models.

​

Bugfixes & improvements

• Fixed a problem in title chat generation that it was creating titles without capital letters.
• Show all tree view for confluence documents when picking files to link
• When doing an interaction documents without content are ignored
• On a existing confluence integration, in case of error during authentication now it tries to log in again
• Force streaming mode in low level API

​

Bugfixes & improvements

• Improved the user experience when a user removes themself from a project, showing a warning.

• Released to Sandbox: July 8, 2025
• Released to Stable: -

​

Bugfixes & improvements

• Optimized how FE cache works
• Added supported for citations in thinking blocks
• Fixed a problem that caused summaries to crash in some scenarios
• Fixed a problem in source selector, that prevented to use Zylon AI even when it was selected

​

Tool selector

​

Agentic Zylon AI

​

Bugfixes & improvements

• Trying to ingest documents from Confluence was failing due to a breaking change in Atlassian API
• Remove file from source selector when the original document has been removed
• Improved chat UI
• Show empty state for file system integration
• Citation with PDF files now open the correct page

​

Agentic Zylon AI - Alpha tester program

​

Transfer project ownership

​

Link files from different projects

• Light files get higher priority than heavy files on ingestion queue to expedite the process and avoid queues blocked by a single heavy file. More info in (Optional) Performance tweaks
• When a non-UTF-8 HTML is ingested, now the content is processed correctly
• Improved CSV delimiter and header auto-detection
• Improved tables ingestion for higher data retrieval accuracy
• In case Claromentis credentials have changed a modal appears to update them

​

File System Knowledge Base Integration

• Changing content of Agent Flows threw an error even though the content was saved properly
• Changing org in different tabs no longer generates flickering
• Navigate to settings without projects
• Code block extension inside the editor has been improved

​

Dashboard analytics for admins

​

Automatic chat name

​

Whitelabel

• Keep your agent’s name under 20 characters.
• Avoid using mixed casing (e.g., “MyAIAssistant”).
• Do not include unusual symbols, special characters, or excessive punctuation (for example, “$uper_Bot!”).

assistantName: "Zylon"

• Qwen 2.5 14B is the new default model in the baseline-24g preset.

​

Integrations owners from Confluence, Claromentis or Sharepoint can sync ingested files and folders

​

Pin your favourite projects for quick access

• UI performance
• Improve readability in long columns in tables inside agent flows
• Added AI preset for 32Gb (Mistral Small 3 2501)
• Supported Blackwell GPUs (RTX 5080, RTX 5090)
• Project organization overview UI has been improved. Now you can easily check for private, shared and pinned projects.
• Sources in BulkQ&A Agent Flows now point to the document reference instead of showing the filename and page.
• Long chats are now supported, previously when there was a long chat, new messages disappeared after Zylon response.
• PPT files are now properly ingested.
• Improved documents retrieval.

​

Changes required before updating to Zylon 1.27.3

1. Update Zylon-CLI:

sudo zylon-cli self-update

1. (Optional) Do a backup using zylon-cli before starting the process.
2. Prepare and update Nvidia Drivers and Toolkit:

sudo zylon-cli setup

1. Reboot machine and check that nvidia-smi is working well:

sudo reboot
nvidia-smi

1. Update Zylon to 1.27.0 using:

sudo zylon-cli update

​

New organization picker to switch organizations without logging out

​

Sharepoint connector

• Observability framework
• Now a tooltip with the complete name is shown when a source name is too long in Knowledge Base.

• Reset custom instructions modal to empty state when it was closed without saving
• The flow to add users manually to an organization has been improved to make it easier.
• Usability improvements in several application flows

• sentry.io (you should have this already properly configured)
• *.grafana.net (this is the new one required)

sentry:
enabled: true
environment: "my-company"

observability:
crashReporting: true
usageMetrics: true

companyIdentifier: "my-company"

observability:
crashReporting: false
usageMetrics: false

​

New Features

• Chats can now be copy-pasted and it will respect format
• The file picker allows multi file upload

​

Bugfixes

• Improved Claromentis integration to account for symlinked files which crashed the integration. Symlinks are not ingested.
• Made Claromentis file pulling more resilient. A failed file won’t crash the ingestion process; the file will be marked as failed and the process will continue.

​

Enhanced AI Editor

• Inline AI: You can now prompt inline within a document to include more information or modify existing text. Simply start typing /zylon.
• Real time collaboration: Co-create with your peers in real time!
• CSV export for tables: Download the results of Bulk Q&A or any other table as a CSV file.
• Improved UX: Enjoy a more flexible table editing experience and more versatile text editing features.

​

Knowledge Base Search

​

EML support

​

CSV Export in Chat