HyEm: Query-Adaptive Hyperbolic Retrieval for Biomedical Ontologies via Euclidean Vector Indexing
New method integrates hyperbolic embeddings into standard databases, solving a key bottleneck for medical RAG systems.
A team of researchers has introduced HyEm, a novel retrieval layer designed to solve a core problem in biomedical AI: effectively grounding language models in complex, hierarchical ontologies. Current production systems for retrieval-augmented generation (RAG) rely on standard Euclidean vector databases, which struggle to represent the deep 'is-a' taxonomies found in medical resources like the Human Phenotype Ontology (HPO) or Disease Ontology (DO). While hyperbolic embeddings are mathematically better suited for hierarchies, they face practical adoption barriers due to a lack of native database support and potential performance drops on non-hierarchical queries.
HyEm elegantly bridges this gap. It learns 'radius-controlled' hyperbolic embeddings from ontology data, then uses a log-map transformation to store them as vectors in conventional Euclidean indexes like FAISS or Pinecone for fast candidate retrieval. A second-stage, exact hyperbolic reranking step then occurs. Crucially, a learned query-adaptive gate dynamically determines the mixing weight between standard semantic similarity and hyperbolic hierarchical distance for the final ranking. According to the arXiv paper, this approach maintains 94-98% of the performance of a pure Euclidean baseline on straightforward entity-centric queries while delivering substantial gains on queries that require navigating ontology hierarchies, all with a manageable computational overhead.
- Solves the hierarchy-aware retrieval problem for biomedical RAG by integrating hyperbolic embeddings into standard Euclidean vector databases.
- Uses a query-adaptive gate to dynamically blend similarity scores, preserving 94-98% of baseline performance on non-hierarchical queries.
- Enables practical deployment without overhauling existing ANN infrastructure, requiring only moderate oversampling during retrieval.
Why It Matters
This unlocks more accurate and reliable AI assistants for doctors and researchers by letting them query complex medical knowledge graphs effectively.