Current

Multi-vector retrieval

Benchmark first

Corpora: BEIR (SciFact, ArguAna, HotpotQA) · LoTTE · KILT

Protocol: Build from corpus documents only; issue the published test queries; preserve ranked IDs and component scores. Report nDCG@10, Recall@100, MRR, p50/p95 query latency, index bytes, and results with ACL pre-filtering. Ablate candidate generation, reranking, fusion, graph expansion, and packing independently.

Mari implements MUVERA fixed-dimensional candidate generation, PolarQuant compression, and exact normalized MaxSim reranking in one retrieval path.

Query token vectors

MUVERA

Candidate documentsallowed IDs only

exact MaxSim

RetrievalHit[]ranked + scored

retrieve.py

from mari_components.retrieval import FDEConfig, build_index, search_index

index = build_index({doc.document_id: token_vectors},
    config=FDEConfig(repetitions=20, projection_dimension=16))
hits = search_index(index, query_token_vectors, limit=8,
    allowed_document_ids=authorized_document_ids)

serialize_index and deserialize_index use versioned, checksummed payloads. exact_maxsim is public for direct scoring.

Authorization must precede scoring. Supply allowed_document_ids; post-filtering can leak information through ranks and fallback behavior.

How it works and backing algorithms

Mari’s current path uses token-level late interaction: each query token takes its maximum similarity to any document token, and the maxima are summed. MUVERA maps those multi-vector sets to fixed-dimensional encodings for fast candidate generation; Mari then reranks the candidates with exact MaxSim. The packed Polar codec is an implementation-level compression of candidate encodings, not an alternative relevance model.

Status	Index family	Representation and algorithm	Appropriate when	Primary source
Current	MUVERA + exact MaxSim	Multi-vector FDE candidate generation, compressed storage, exact late-interaction reranking	Fine-grained semantic matching where individual query terms matter	MUVERA · ColBERT
Current	Dense flat	Exact cosine, dot-product, or L2 scan over one vector per passage	Small corpora, evaluation baselines, or exact reproducibility	Dense Passage Retrieval
Current	HNSW	Deterministic hierarchical proximity graph with configurable search breadth	Approximate dense search and recall-versus-exact evaluation	HNSW
Current	IVF-PQ	Coarse inverted partitions plus residual product-quantized vector codes	Memory-constrained dense indexes	Product Quantization · Faiss
Current	BM25	Robertson–Walker lexical ranking over an in-memory inverted representation	Exact names, identifiers, code symbols, and domain terminology	BM25 and Beyond
Current	Learned sparse	Exact sparse inner product over caller-produced term weights	Model-neutral serving for SPLADE-like expansion vectors	SPLADE
Current	Rank fusion	Weighted reciprocal-rank fusion over independent result lists, with per-source contribution traces	Mixed corpora where source scores are not directly comparable	RAG-Fusion
Current	Graph propagation	Allowed-node personalized PageRank followed by weighted node-to-passage projection	Multi-hop recall from query-linked entities, facts, or sections	HippoRAG

Index interfaces

The index classes expose the same authorization-aware search shape while retaining algorithm-specific controls. Index selection is pipeline configuration and can be evaluated against recall, latency, memory, freshness, and ACL-filter behavior.

indexes.py

from mari_components.retrieval import (
    BM25Index, DenseFlatIndex, HNSWIndex, IVFPQIndex, SparseVectorIndex,
)

exact = DenseFlatIndex(vectors, metric="cosine")
graph = HNSWIndex(vectors, metric="cosine", m=32)
compressed = IVFPQIndex(vectors, partitions=256, subquantizers=16)
lexical = BM25Index(passages, k1=1.2, b=0.75)
sparse = SparseVectorIndex(model_generated_term_weights)

hits = graph.search(query_vector, limit=20, ef_search=128,
    allowed_document_ids=authorized_document_ids)

Rank fusion, graph recall, and diverse packing

MUVERAlexicalrecent

RRF

authorized nodesPageRankpassage projection

MMR

Context candidatesscores and contributions retained

compose_retrieval.py

from mari_components.retrieval import (
    maximal_marginal_relevance, personalized_pagerank,
    project_graph_scores, reciprocal_rank_fusion,
)

fused = reciprocal_rank_fusion(
    {"muvera": dense_ids, "lexical": lexical_ids, "recent": recent_ids},
    weights={"recent": 0.25}, rank_constant=60,
    eligible=authorized_document_ids.__contains__, limit=40)

nodes = personalized_pagerank(graph, query_seeds,
    allowed_node_ids=authorized_graph_nodes, damping=0.85)
passages = project_graph_scores(nodes.hits, node_passages, limit=20)

context = maximal_marginal_relevance(
    {hit.document_id: hit.score for hit in fused},
    similarity=passage_similarity, relevance_weight=0.65, limit=12)
assert nodes.converged