Today's work focused on enhancing the AI Fill experience and addressing performance bottlenecks. A major user-facing update is that the "Fill with AI" button for individual fields now creates a suggestion card, mirroring the behavior of the bulk fill feature. This allows users to review and accept AI-generated values before they are applied, providing more control and transparency. We also added a confidence indicator to each AI fill suggestion card, providing an at-a-glance assessment of the model's certainty. This helps users quickly evaluate the reliability of the suggestion. Finally, we added new developer actions to reset report data and files, speeding up development workflows.

Significant effort was dedicated to improving the reliability and performance of AI Fill, especially when using Anthropic's Claude models. We migrated to AI SDK v6, updated default models, and implemented several fixes to address rate limiting and schema validation issues. The AI fill functionality now leverages the search_data RAG tool and defaults to using Grok, optimizing for speed and cost while staying within token limits. We also implemented stricter tool ordering and output limits to control token usage during the AI agent's exploration.

We also fixed a bug where Office lock files could cause ingestion to fail. The system now rejects these files and strips invalid characters from parsed content. Finally, we made several fixes related to error surfacing in AI fill.