rerank_bridge_mmr
rerank_bridge_mmr(
query_emb,
candidate_indices,
embeddings_normed,
bridge_scores,
cluster_labels,
top_k,
lam=0.5,
meta_clusters=None,
)MMR with bridge-weighted diversity and optional meta-cluster awareness.
Instead of max cosine similarity to selected set, penalizes candidates that share a cluster with already-selected results. Bridge points get a diversity bonus because they span clusters.
If meta_clusters is provided (set of cluster IDs), those clusters don’t count as “new coverage” — the algorithm won’t waste diversity slots chasing structural/meta clusters (dates, years, event lists).
Parameters
| Name | Type | Description | Default |
|---|---|---|---|
| query_emb | (d,) normalized query embedding. | required | |
| candidate_indices | 1-D array of candidate point indices. | required | |
| embeddings_normed | (n, d) L2-normalized embedding matrix. | required | |
| bridge_scores | Array of bridge scores indexed by point index. | required | |
| cluster_labels | Array of cluster labels indexed by point index. | required | |
| top_k | Number of results to return. | required | |
| lam | Relevance-diversity tradeoff. | 0.5 |
|
| meta_clusters | Optional set of cluster IDs to suppress for diversity credit. | None |
Returns
| Name | Type | Description |
|---|---|---|
| np.ndarray of selected point indices. |