rankweave

Language-agnostic hybrid-retrieval score fusion — pure-Python, store-agnostic.

rankweave decides how to combine the scores from a lexical channel (character-trigram / BM25 / learned-sparse) and a semantic channel (dense embeddings) into one ranking. It ships two research-grounded fusion strategies and the query-side Unicode normalization that makes character-level lexical matching language-agnostic. It has no dependencies (stdlib only) and no opinion about your store — bring your own channels; rankweave fuses their scores.

It is extracted, unchanged in behavior, from the Context Search engine of naruon, following the lab's ONE SOURCE MULTI USE convention: standalone product and submodule-importable.

Why

A convex combination of theoretically min-max normalized scores (TM2C2) beats Reciprocal Rank Fusion in- and out-of-domain, is robust for alpha ∈ [0.6, 0.8] with no training data, and — unlike rank fusion — preserves the score distribution (Bruch, Gai & Ingber 2023). RRF remains available for channels that expose only ranks. See docs/research/ for the grounding.

Install

pip install rankweave

Quickstart

from rankweave import FusionSettings, fuse_channel_scores, normalize_search_text

# 1) Normalize the query the same way you normalize indexed documents
#    (NFC compose; do accent-folding + lowercasing on the store side too).
query = normalize_search_text("  Trần Hưng Đạo 회의  ")   # -> "Trần Hưng Đạo 회의"

# 2) Run your own lexical + dense channels, then fuse per candidate.
settings = FusionSettings()                 # TM2C2, semantic weight alpha = 0.7
score = fuse_channel_scores(
    word_similarity_score=0.62,             # lexical channel score in [0, 1]
    cosine_distance=0.30,                   # dense channel distance in [0, 2]
    channel_ranks={"lexical": 1, "dense": 1},
    settings=settings,
)                                            # -> bounded [0, 1] fused score

A channel that did not return a candidate contributes its theoretical minimum (absent evidence is the infimum, not an imputed value). Pass FusionSettings(strategy_name="reciprocal_rank_fusion") to fuse by rank instead — then only channel_ranks matters.

API

The normalization contract

Character-trigram lexical retrieval is language-agnostic only if query and documents fold identically. normalize_search_text owns the query side (NFC). Do accent-folding + lowercasing on the store side, in one place, and call it from both — e.g. a PostgreSQL IMMUTABLE wrapper lower(unaccent(normalize(text, NFC))) used in a pg_trgm GiST expression index. rankweave stays out of your store so the two sides cannot silently diverge.

Research grounding

• Bruch, Gai & Ingber (2023). An Analysis of Fusion Functions for Hybrid Retrieval. ACM TOIS 42(1). arXiv:2210.11934. — TM2C2 > RRF; theoretical-normalization stability; the fusion desiderata (monotonicity, homogeneity, boundedness, Lipschitz continuity, sample efficiency) this library's defaults satisfy.
• Cormack, Clarke & Büttcher (2009). Reciprocal Rank Fusion outperforms Condorcet and individual Rank Learning Methods. SIGIR 2009. — RRF definition, η = 60.
• UAX #15, Unicode Normalization Forms — NFC composition.

PDFs and a citation manifest live in docs/research/.

Development

pip install -e ".[dev]"
python -m pytest -q          # 27 tests, no external services
python -m ruff check .

License

Apache-2.0 — see LICENSE.