hyper³labs

Embedding models and retrieval infrastructure.

Website · GitHub

hyper³labs (pronounced Hypercube Labs) builds embedding models and retrieval infrastructure across text and vision. We also build open-source tools for evaluating and understanding embedding systems.

We are especially interested in how hierarchy and non-Euclidean geometry affect retrieval, and in the failures that aggregate benchmark scores hide.

Models

• hyper3-clip-v0.5 is an open-weight vision-language embedding model for hierarchy-sensitive retrieval. It produces 512-dimensional embeddings and works with the Sentence Transformers interface.

Tools and demos

• HyperView is our open-source workbench for exploring embedding spaces and tracing retrieval failures back to real samples.
• ABO Catalog Explorer tests fine-grained retrieval on product catalogs.
• DeepFashion Search explores text-to-image search for garments and product attributes.
• Precision Region Search demonstrates referring-expression retrieval on RefCOCOg.
• Jaguar Re-ID examines whether re-identification models recognize an animal or its background.
• VisA Industrial Search explores anomaly and part retrieval in industrial imagery.

Datasets

• jaguar-re-id: individual jaguar identification from spot patterns.
• amazon-berkeley-objects: a mirror used for hierarchical product retrieval evaluation.
• jaguar-hyperview-demo: a smaller subset prepared for interactive analysis.

Research

Are We Recognizing the Jaguar or Its Background? A Diagnostic Framework for Jaguar Re-Identification
Antonio Rueda-Toicen, Robert Vava, Matin Mahmood. April 2026.