Changelog

Deep Research Search Guardrails

• Deep research prompts now auto-trigger web search whenever the knowledge base lacks a direct answer, so clinicians get complete guidance without manual toggles.
• Web search is now constrained to vetted, authoritative clinical domains to keep retrieved evidence trustworthy and citeable.
• Rename "Knowledge Base" model to "Clinical Guidelines" to better reflect its focus on guideline-based answers.
• Rename "Deep Research" model to "Guidelines + Web Search" to clarify its combined sourcing approach.
• Update model descriptions to highlight the emphasis on guidelines and trusted web sources for improved clarity.

GPT-5.1 Upgrade

• Swapped GPT-4.1 for GPT-5.1 across chat and document creation, and tightened model registry tests so provider routing stays explicit.
• Fingerprinted every system prompt and dedupe prompt versions so telemetry, analytics, and audit trails line up even when prompts are reused in different workflows.

Document creation mode routing

• Added smarter intent detection so the assistant knows when you want a quick answer versus a drafted letter or note.
• Updated the chat pipeline to feed document-style requests the right prompts, models, and safety steps without extra clicks.
• Shipped new helper code and tests that keep the document drafting workflow stable as we tweak prompts over time.

Switch Drug Dosing Detection Model

• Drug dosing detection model updated to grok-4-fast-reasoning from gpt-5-mini for improved accuracy.

Changelog page goes live

• Launch the /changelog page so updates are easy to follow.
• Add the first entries covering the beta release and feature highlights.
• Surface an “Intended Use” link in the sidebar so clinicians can quickly review scope and guardrails.

Voice input arrives

• Delivered speech recognition support inside the multimodal input so clinicians can dictate questions hands-free.
• Enhanced the MultimodalInput component to streamline recording controls and added tests covering the new voice flow.
• Updated deep research prompts to make sure voice-captured queries receive the same high-quality guidance as typed requests.

Prompt refresh for clarity

• Updated the assistant prompts for clearer guidance and better accuracy during clinical conversations.
• Refined deep research instructions so long-form answers stay concise while covering critical safety context.

UX polish and content tooling

• Modularized database queries and refreshed documentation across chat, export, guideline, and profile modules to keep contributors oriented.
• Added Open Graph image generation plus sitemap and robots endpoints so shared links display rich previews and search engines stay informed.
• Improved chat responsiveness with a Spinner component and refined model selector integration inside the multimodal input.
• Tuned message handling, reorganized imports, and shipped supporting tests to keep the UI snappy during heavy usage.

ChatIBD Beta Launch

ChatIBD Beta launch introducing the AI assistant for IBD clinicians and the structured dosing cards that anchor guideline-backed responses.

• Implemented shared caching for guideline fetches so repeated clinical lookups respond faster without redundant upstream calls.