KeySI

An interaction framework for tuning text embeddings based on human feedback.

KeySI enables feature-level feedback for embedding model refinement. Users organize extracted keywords into concept groups, which the system translates into document-level supervision for subsequent tuning—lowering the barrier to adapting embeddings without large labeled datasets.

I contributed to study design, pilot sessions, and interface evaluation as second author.

Outcomes

Collaborators: Yan Zhu, Prof. Rebecca Faust