gardian-reranker 0.1.0

Query-adaptive sparse/dense/KG reranker for RAG systems.

gardian-reranker

Query-adaptive sparse/dense/KG reranker for RAG systems.

Install

pip install .

After publishing to PyPI:

pip install gardian-reranker

What it returns

gardian-reranker returns:

• query-level adaptive branch weights: sparse/dense/kg
• ranked passages with final score
• per-passage branch contributions for explainable routing into the reader

Python API

from gardian_reranker import GardianReranker, Candidate, QueryFeatures

reranker = GardianReranker.from_checkpoint("gardian_best.pt", {
    "sparse_dim": 3,
    "dense_dim": 4,
    "kg_dim": 6,
    "branch_hidden": 256,
    "controller_hidden": 256,
    "query_feat_dim": 1024,
    "n_qtypes": 7,
    "dropout": 0.2
})

query_features = QueryFeatures(
    query_emb=[0.0] * 1024,
    qtype_onehot=[0, 0, 0, 0, 0, 0, 1],
    kg_coverage=0.8
)

candidates = [
    Candidate(
        id="doc1",
        text="Passage text",
        sparse_feats=[0.1, 0.2, 0.3],
        dense_feats=[0.4, 0.5, 0.6, 0.7],
        kg_feats=[0.1, 0.2, 0.3, 0.1, 0.0, 0.4],
    )
]

result = reranker.rerank(query_features, candidates)
print(result["query_weights"])
print(result["ranked"][0]["contrib"])

CLI

gardian-rerank \
  --checkpoint results/gardian_best_hybrid.pt \
  --config config.json \
  --input input.json \
  --output output.json

PyPI release (maintainers)

1. Replace placeholder repository URLs in pyproject.toml.
2. Configure Trusted Publishing for your PyPI project.
3. Bump version in pyproject.toml.
4. Push a tag, e.g. v0.1.1:

git tag v0.1.1
git push origin v0.1.1

The GitHub Action .github/workflows/publish-pypi.yml will build and publish automatically.