We've significantly improved the AI Fill feature by implementing deduplication guards for retrieval tools. This prevents redundant searches and file fetches, saving tokens and improving efficiency. We also introduced a hybrid retrieval context, enhancing the relevance and accuracy of AI-generated suggestions by incorporating field hints and optimizing spreadsheet handling. These improvements collectively reduce wasted tokens, streamline the AI's processing, and enhance the overall quality of AI-driven insights. Additionally, we've added a safeguard against runaway tool loops by implementing a hard budget for single-field and bulk AI fill, ensuring stability and preventing the exhaustion of resources. This enhancement ensures more predictable and reliable performance for our users. Finally, the 'View file' button in report Sources is now functional, allowing users to directly access and view the source files referenced in reports. This streamlines the process of verifying information and enhances user trust in the platform's data.

We focused on refining the user experience when interacting with repeater fields in chat. Suggestions are now rendered as styled tables, providing a clear and intuitive way to visualize data. The opening message in the chat dialog also presents the current repeater value in a user-friendly format. We've also normalized repeater suggestions to use column IDs, ensuring seamless data integration when accepting suggestions from the chat-about-field dialog. These enhancements provide a more polished and intuitive user experience when working with complex data structures.

Several fixes addressed correctness bugs and UI consistency issues. We ensured that duplicate calls do not consume the per-field budget and that failed searches do not poison the deduplication cache. Furthermore, we've enhanced the UI with semantic shadcn tokens for improved dark-mode support and added a focus-visible ring for keyboard users. These meticulous refinements ensure a robust and visually consistent user experience.

Finally, we've made several improvements to the underlying infrastructure to ensure data security and efficient processing. We've enforced scoped retrieval safeguards, sanitized retrieval error logging, and addressed potential prompt-injection risks. Additionally, we've optimized file chunk processing and implemented checks to prevent cross-tenant rechunking. These under-the-hood improvements strengthen the platform's reliability and security posture.