small-hybrid-reranker

small-hybrid-reranker is a lightweight reranker package with a baked-in trained model.

It reranks a list of passages for a query using a hybrid feature stack:

• static embeddings (cnmoro/static-nomic-384-pten-v2)
• lexical overlap and token interaction sketches
• BM25 and dense retrieval priors
• listwise LightGBM ranker

The model artifact is included in the package, so there is no separate checkpoint download.

Model In This Release

• Version 0.2.1 packages a model retrained with cnmoro/static-nomic-384-pten-v2.
• Training used the SciFact train split in this repository with strict BM25 top-100 candidates and LightGBM LambdaRank over hybrid features.
• In-split training metric from the full training run:
  • ndcg@10: 0.90908
  • recall@10: 0.90766

Inference remains lightweight and CPU-friendly: the API is still a single HybridReranker().rerank(query, passages) call.

Install

pip install small-hybrid-reranker

Quickstart

from small_hybrid_reranker import HybridReranker

reranker = HybridReranker()

query = "What is the speed of light?"
passages = [
    "The speed of light in a vacuum is about 299,792 km/s.",
    "Earth orbits the Sun in about 365 days.",
    "Newton described laws of motion.",
]

ranked = reranker.rerank(query, passages)
print(ranked[0])
# {'passage': 'The speed of light in a vacuum is about 299,792 km/s.', 'score': 100.0}

API

HybridReranker(model_path: str | None = None)

• model_path=None: uses baked-in model inside package.
• model_path="...joblib": load your own compatible artifact.

rerank(query: str, passages: list[str], top_k: int | None = None) -> list[dict]

Returns:

[
  {"passage": "...", "score": 82.31},
  {"passage": "...", "score": 40.87},
]

Scores are floats in [0, 100] and sorted descending.

Notes

• This package is optimized for reranking a provided candidate list.
• It is not a full retrieval system by itself.