Changelog

2.9 - Apr 22, 2026
new multi-phase sources pipeline, user-controlled source configuration, geolocation for local guidelines, and reasoning transparency

• Redesigned sources pipeline: Introduced a new multi-phase retrieval architecture that separates source gathering into distinct phases, improving the relevance and structure of information used to generate responses.

  
  

• Non-blocking parallel source generation: Sources are now fetched concurrently and non-blocking, reducing wait times and making responses feel significantly faster.
  
  

• User-controlled source configuration: Users can now configure and persist their preferred source types directly from a reworked sources panel. Source preferences are saved to their account and applied consistently across sessions.
  
  

• Geolocation for local guidelines: Added geolocation detection to automatically surface locally relevant clinical guidelines. Users can set or override their location directly from the sources panel.
  
  

• Merged and deduplicated references: References are now grouped and merged by source type and label, reducing noise and making the citations panel easier to scan.
  
  

• Reasoning transparency: Thinking tokens from the pipeline are now surfaced in the UI and saved to the database, giving users visibility into the model's reasoning process as it works through a query.
  
  

• Citation improvements: Improved fallback handling when no source is extracted, and a bug with follow-up queries on web sources has been fixed.
  
  

• General fixes and polish: Resolved issues with source selection not carrying over on edited messages, numbered list clipping, scrollbar in markdown tables, session file content loss, and various stability improvements.

• Significantly improved answer accuracy: Major upgrades to the underlying reasoning and retrieval pipeline to provide more precise and reliable responses to user queries.

• Expanded and enhanced source coverage: Increased both the number and quality of sources used to generate answers, improving the depth and relevance of retrieved information.

• Greater user control over sources: Users can now influence which source types are prioritised and how information sources are weighted when generating responses.

• General user experience improvements: Various interface and performance refinements to make the application smoother, faster, and easier to use.

• New “Information Gathering Mode”: Introduced a toggleable mode that suppresses AI responses when messages are sent. This allows users to continuously add transcripts or supplementary information without interruption before requesting a response.

• Auditing feature dialogue made public

• Wide ranging usability improvements and bug fixes

• Evidence linking for spaces. Like in coding, the evidence for a particular statement is highlighted in yellow when you click on it.