NASA’s new dark energy space telescope can also detect killer asteroids
technologyreview.com·1h ago
AI Research
arXiv CS.AI·May 21·6 min read
YAC: Bridging Natural Language and Interactive Visual Exploration with Generative AI for Biomedical Data Discovery
TL;DR
Biomedical data discovery tools typically force users to choose between natural language queries or interactive visual exploration—not both. YAC (Yet Another Chatbot) combines these by using a multi-agent system with tool-calling (where an AI decides which functions to invoke) to convert natural language into declarative visualizations that users can then refine interactively.
✦ Why It Matters
Engineers building data discovery interfaces can adopt multi-agent tool-calling patterns to let users query complex datasets conversationally while maintaining interactive control.
Key Takeaways
- Use multi-agent systems with tool-calling to bridge natural language and structured visualization specifications.
- Implement declarative output generation to decouple user intent from visualization rendering logic.
- Combine conversational AI with interactive visual exploration rather than choosing one modality.
- Apply this pattern to domain-specific data discovery where users need both query expressiveness and visual feedback.
Full Summary
Biomedical researchers need efficient ways to explore complex datasets, but existing interfaces force a choice between natural language queries (conversational but limited in precision) and interactive visualizations (powerful but require manual navigation). YAC addresses this gap by building a multi-agent system where language models coordinate tool calls—function invocations that trigger specific actions—to generate declarative output (structured specifications rather than imperative commands).
The system interprets this output to render interactive visualizations, allowing users to query data conversationally while maintaining visual exploration capabilities. The architecture uses generative AI (large language models) to bridge the semantic gap between natural language intent and visualization specifications.
This approach preserves the strengths of both interaction modalities: users can express complex queries naturally while retaining visual feedback and interactive refinement. The prototype demonstrates feasibility for biomedical data discovery workflows, though specific quantitative results are not detailed in the abstract.