Documentation
Definitions and methods behind Genefold’s spectral approach to embedding retrieval and monitoring. Each page opens with a direct definition, states its assumptions, and links to the primary evidence.
Spectral intelligence
Using eigenvalues, eigenvectors, and graph operators to describe structure in embedding spaces for retrieval and monitoring.
MethodSpectral vector search
Ranking embeddings with graph structure and bounded spectral scores, combined with cosine similarity at query time.
MethodGraph Laplacian retrieval
The mathematical formulation: feature-space graphs, Rayleigh quotients, graph wiring, and bounded spectral scores.
OperationsEmbedding drift
Detecting distribution shift in embedding streams, and what spectral signals add to embedding-level statistics.
MonitoringOut-of-distribution vector retrieval
How queries outside the corpus manifold behave, and which diagnostics expose them before they mislead retrieval.
ComparisonSpectral search vs HNSW
A qualitative comparison of signals, index structures, long-tail behaviour, and build costs against standard ANN indexes.
ReferenceGlossary
Canonical, short definitions of the core terms used across Genefold documentation and papers.