This release brings a major step forward in answer quality with Re-ranking, plus a batch of improvements across Knowledge Base web content, MCP tools, navigation, integrations, and upload APIs.
Re-ranking for Better Answers
Re-ranking adds a second, smarter selection step after your initial search (vector, keyword, or hybrid). It improves answer quality by ensuring the AI uses the best knowledge, not just the closest embeddings.
How it works
-
Step 1: Retrieval
The system retrieves a set of “probably relevant” candidates
Example: top 20 chunks from a vector database or hybrid search results. -
Step 2: Re-ranking
A more precise model evaluates each candidate against the real user question and then:-
assigns a relevance score
-
sorts results from most to least relevant
-
optionally discards weak matches
-
Why it matters
Vector similarity is powerful, but it can:
-
return chunks that are semantically similar but contextually wrong
-
treat all retrieved candidates as equally good once they are in the shortlist
-
miss intent, constraints, and priorities in the user request
Re-ranking solves this by comparing the user question directly with each candidate chunk, resulting in more accurate answers and higher confidence in the retrieved sources.
Knowledge Base Web Content Improvements
We improved how Knowledge Base web sources are managed and maintained, especially for teams ingesting documentation sites.
-
Synchronized sitemaps
Knowledge Bases can now add synchronized sitemaps, making ongoing website ingestion more structured and maintainable. -
Simplified indexing options for web imports
Removed the “Standard” indexing option from Advanced options in:-
Import from Websites
-
Import from Sitemap
This reduces confusion and keeps web content ingestion aligned with the current recommended approach.
-
-
Copy and paste scraping rules
You can now copy scraping rules from one web content source and paste them into a new one.
This makes it much faster to configure multiple sources consistently. -
Bug fix: KB selection with “use KB name”
Fixed an issue where Knowledge Base selection could break when “use KB name” was enabled.
MCP Servers and Tooling Updates
We kept building on MCP by expanding tools and making tool management smoother.
-
Adds MCP server tools
More tools are available via MCP servers, enabling richer task execution from your AI Agents. -
Improved saving and refresh behavior for MCP tools
Tool saving and refreshing is now handled more reliably, reducing friction during setup and iteration. -
UI refinements for MCP tool selection
-
Updated badge colors in the model list
-
Cursor behavior improved on selected tools
-
Updated label from “Chose…” to “Manage MCP Tools” for clarity
-
Navigation and Dashboard UX
Small workflow improvements that save time every day:
-
History to conversation details navigation improved
When you open a conversation from History and go back, the dashboard now remembers the starting list page so you do not lose your place. -
Dashboard language selection on “Your Projects” page
You can now choose your dashboard language directly from the Your Projects page.
Model Support Updates
-
Added Anthropic Claude Opus 3.6 to supported LLMs.
Upload Service and API Improvements
We updated the upload system to improve consistency and future extensibility:
-
Updated the file upload service
-
Updated APIs for uploading into chat:
-
API for upload a file/image into chat
-
API for upload a file/images
-
Sign-in and Platform Integrations
-
Sign in with Google is now available.
-
Updated Launch.js for Wix and Shopify by bypassing
srcdoc
This improves compatibility and reliability in those environments.
Flow Builder Expansion
-
Added Iteration action
This enables repeatable logic in flows, useful for structured, step-by-step automation patterns.
With Re-ranking and MCP improvements, your AI Agent moves closer to true task automation, using better knowledge, smarter selection, and real tool execution.
