Crate sphereql_embed 

Vector embedding projection engine.

Projects high-dimensional embeddings onto S² via one of several projection families (PCA, kernel PCA, Laplacian eigenmap, UMAP) and offers a query pipeline for k-NN search, similarity thresholds, concept paths, glob detection, local manifold fitting, and a category-level enrichment layer (inter-category graph, bridge detection with Genuine/OverlapArtifact/Weak classification, inner spheres, hierarchical concept paths, domain-group routing for low-EVR regimes).

On top of that, the crate ships a metalearning framework:

• config — a single PipelineConfig hierarchy for every tunable constant.
• quality_metric — a QualityMetric trait plus four concrete metrics (territorial health, bridge coherence, cluster silhouette, graph modularity) and composite presets.
• tuner — auto_tune over a discrete SearchSpace with Grid / Random / Bayesian (TPE-lite) strategies. Projection kind is a first-class tuner axis.
• corpus_features — a 10-feature corpus profile suitable as input to a meta-model.
• meta_model — MetaTrainingRecord with an on-disk store under ~/.sphereql/meta_records.json, the MetaModel trait, and two concrete implementations (NearestNeighborMetaModel, DistanceWeightedMetaModel).
• feedback — per-query user-satisfaction primitives (FeedbackEvent + FeedbackAggregator) for L3 online refinement of stored records.

See SphereQLPipeline::new_with_config, SphereQLPipeline::new_from_metamodel, and SphereQLPipeline::new_from_metamodel_tuned for the tune-or-recall entry points.

Re-exports

pub use umap::UmapSphereProjection;

pub use category::*;

pub use confidence::*;

pub use config::*;

pub use configured_projection::*;

pub use corpus_features::*;

pub use corpus_quality::*;

pub use domain_groups::*;

pub use feedback::*;

pub use kernel_pca::*;

pub use laplacian::*;

pub use logical_confidence::*;

pub use mapper::*;

pub use meta_model::*;

pub use navigator::*;

pub use pipeline::*;

pub use projection::*;

pub use quality_metric::*;

pub use query::*;

pub use self_tune::*;

pub use spatial_quality::*;

pub use text_embedder::*;

pub use tuner::*;

pub use types::*;

pub use util::*;

Modules

ann

Approximate nearest neighbors via random projection forest.

category

confidence

Unified confidence scoring for query results.

config

Configuration surface for the SphereQL pipeline.

configured_projection

ConfiguredProjection — a single concrete type over all supported outer-sphere projection families.

corpus_features

Low-dimensional profile of a corpus — features that meta-learning can map to an optimal PipelineConfig.

corpus_quality

Composite corpus-quality metric.

domain_groups

Hierarchical domain groups: coarse routing for noisy projections.

feedback

User-supplied feedback signals that refine stored meta-learning records.

kernel_pca

laplacian

Laplacian eigenmap projection — connectivity-preserving embedding on S².

logical_confidence

Skeleton for projection-aware logical-confidence scoring.

mapper

meta_model

Meta-learning across corpora: predict a PipelineConfig for a new corpus by consulting past tuner runs on similar corpora.

navigator

AI Knowledge Navigator — semantic spatial queries on S².

pipeline

projection

quality_metric

Pluggable quality metrics for auto-tuning a SphereQL pipeline.

query

self_tune

Post-hoc corpus self-tuning algorithm.

spatial_quality

Spatial quality metrics computed from category geometry on S².

text_embedder

Pluggable text-to-embedding hook.

tuner

Auto-tuner: search PipelineConfig space to maximize a QualityMetric.

types

umap

UMAP-on-sphere via Adam in the tangent bundle of S².

util

Tiny shared helpers used across multiple modules.