find_super_connectors

find_super_connectors(
    embeddings,
    global_num_bits=None,
    facet_num_bits=None,
    dense_percentile=75,
    global_threshold_percentile=50,
    local_threshold_percentile=50,
    min_bucket_size=20,
    seed=42,
    bridge_percentile=10,
)

Find super connectors: points with high centrality in both global and local bridge networks.

Super connectors are ideal RAG anchor points because they: - Bridge across major semantic regions (high global centrality) - Connect facets within dense clusters (high local centrality) - Provide 10x better coverage efficiency than random anchors

Parameters

Name	Type	Description	Default
embeddings	np.ndarray	Normalized embeddings (n_points, dim)	required
global_num_bits	int | None	LSH bits for global bucketing. None (default) derives it from the corpus size via :func:_derive_num_bits, which is required for the dense-bucket gate to open at all below ~8k points — the previous fixed default of 12 returned zero super connectors on every smaller corpus.	None
facet_num_bits	int | None	LSH bits for facet bucketing within a dense bucket. None (default) uses two bits below the global resolution.	None
dense_percentile	float	Percentile threshold for dense buckets	75
global_threshold_percentile	float	Percentile for “high” global centrality	50
local_threshold_percentile	float	Percentile for “high” local centrality	50
min_bucket_size	int	Minimum bucket size for faceting	20
seed	int	Random seed	42
bridge_percentile	float	Bridges are the lowest-centroid-similarity points; this is the percentile of this corpus’s similarity distribution used as the cut. Relative by design — analyze_bridges’s absolute 0.5 default returns zero bridges on unit-norm text embeddings (0.0% fall below it) and floods on isotropic data (92.3% below), so an absolute constant cannot serve both.	10

Returns

Name	Type	Description
	SuperConnectorResult	SuperConnectorResult with indices and centrality data