refine_dyf_tree

refine_dyf_tree(
    tree,
    embeddings,
    min_coherence=None,
    num_bits=3,
    min_leaf_size=4,
    max_retries=3,
    seed_offset=1000,
)

Re-split incoherent tree leaves with rotated LSH seeds.

Walks the tree and finds leaves whose mean cosine similarity to centroid is below a threshold. For each, tries re-splitting with different LSH seeds and accepts the split if weighted child coherence improves.

Modifies the tree in-place.

Parameters

Name	Type	Description	Default
tree		DYF tree dict from build_dyf_tree().	required
embeddings		(n, d) array of embedding vectors.	required
min_coherence		Coherence threshold. Leaves below this are candidates for re-splitting. If None, uses 25th percentile of leaf coherences.	None
num_bits		LSH bits for re-splitting (default 3).	3
min_leaf_size		Don’t try to split leaves smaller than 2x this.	4
max_retries		Number of different seeds to try per leaf.	3
seed_offset		Base seed for re-splitting (offset from original tree seed).	1000

Returns

Name	Type	Description
		dict with stats: n_refined, n_leaves_before, n_leaves_after,
		coherence_before, coherence_after.