Skip to main content

Reranking

vectlite includes built-in rerankers that re-score search results after the initial retrieval phase. Rerankers can improve relevance by incorporating signals beyond vector similarity.

info

Rerankers are currently available in the Python binding only.

Text Match Reranker​

Boosts results whose metadata text overlaps with the query:

from vectlite import rerankers

results = db.search(
    query_embedding,
    k=10,
    sparse=vectlite.sparse_terms("SSO authentication"),
    rerank=rerankers.text_match(),
)

Options​

rerankers.text_match(
    text_key="text",           # Metadata field with document body
    title_key="title",         # Metadata field with document title
    text_weight=1.0,           # Weight for body text overlap
    title_weight=1.5,          # Weight for title text overlap
    matched_term_weight=0.25,  # Weight for BM25 matched terms
    phrase_boost=1.0,          # Boost for exact phrase matches
)

Metadata Boost Reranker​

Boosts results based on a metadata field value:

results = db.search(
    query_embedding,
    k=10,
    rerank=rerankers.metadata_boost("source", {"docs": 0.5, "blog": 0.2}),
)

Results with source: "docs" get +0.5 added to their score, source: "blog" gets +0.2.

Cross-Encoder Reranker​

Uses a cross-encoder model for high-quality re-scoring. Requires sentence-transformers:

pip install sentence-transformers
results = db.search(
    query_embedding,
    k=10,
    rerank=rerankers.cross_encoder("cross-encoder/ms-marco-MiniLM-L-6-v2"),
)

Options​

rerankers.cross_encoder(
    model_name_or_path="cross-encoder/ms-marco-MiniLM-L-6-v2",
    text_key="text",     # Metadata field with document text
    batch_size=32,       # Batch size for inference
    device=None,         # "cpu", "cuda", etc.
)

Bi-Encoder Reranker​

Uses a bi-encoder model to compute embedding similarity between query and document text:

results = db.search(
    query_embedding,
    k=10,
    rerank=rerankers.bi_encoder("sentence-transformers/all-MiniLM-L6-v2"),
)

Composing Rerankers​

Chain multiple rerankers sequentially or with reciprocal rank fusion:

Sequential​

results = db.search(
    query_embedding,
    k=10,
    rerank=rerankers.compose(
        rerankers.text_match(),
        rerankers.metadata_boost("source", {"docs": 0.5}),
    ),
)

RRF Composition​

results = db.search(
    query_embedding,
    k=10,
    rerank=rerankers.compose(
        rerankers.text_match(),
        rerankers.cross_encoder(),
        strategy="rrf",
        rank_constant=60,
    ),
)

Rerank with Oversampling​

Fetch more candidates than needed, rerank, then return the top k:

results = db.search(
    query_embedding,
    k=10,
    rerank=rerankers.cross_encoder(),
    rerank_k=50,  # Fetch 50 candidates, rerank, return top 10
)

Explain Mode​

See detailed reranking traces with explain=True:

results = db.search(
    query_embedding,
    k=10,
    explain=True,
    rerank=rerankers.text_match(),
)

for r in results:
    print(r["explain"]["rerankers"])
    # [{"name": "text_match", "base_score": 0.85, "body_overlap": 0.6, ...}]