STEER

Semantic Transformation for Embedding-space Exploration in Retrieval

A retrieval primitive for controllable semantic navigation in embedding spaces. No retraining, no fine-tuning, works with any embedding model.

Paper: STEER: An Honest Evaluation of Sixteen Algebraic Operations

Install

pip install steer-retrieval

Quick Start

from steer import Steerer

# Initialize with any sentence-transformer model
steerer = Steerer(model_name="all-MiniLM-L6-v2")

# Index your corpus
steerer.index(["doc1 text...", "doc2 text...", ...])

# Baseline search
results = steerer.search("VEGF inhibition in retinal disease")

# Steered search: explore adjacent domain
results = steerer.search(
    "VEGF inhibition in retinal disease",
    target="oncology drug mechanisms",
    alpha=0.1,
)

# Contrastive: push toward positive, away from negative
results = steerer.contrastive_search(
    "VEGF inhibition in retinal disease",
    positive_target="oncology drug mechanisms",
    negative_target="basic laboratory research",
    alpha=0.1,
    beta=0.2,
)

# Gradient walk: see how results change across alpha values
walk = steerer.gradient_walk(
    "VEGF inhibition in retinal disease",
    target="oncology drug mechanisms",
)
for step in walk:
    print(f"alpha={step['alpha']}: {step['results'][0]['id']}")

Operations

STEER provides 16 operations built from a single primitive: normalize(q + alpha * t).

Key Findings

• Contrastive asymmetric (alpha=0.1, beta=0.2) produces the strongest results: +3.5% nDCG
• Rotate Away unexpectedly improves retrieval: +3.6% nDCG (bias correction)
• Adaptive alpha reduces degradation by 88-93%
• Multi-vector RRF eliminates 100% of degradation
• Classifier with per-query LLM targets achieves F1=0.91 as automatic router

Citation

@article{gomes2026steer,
  title={STEER: Semantic Transformation for Embedding-space Exploration in Retrieval --- An Honest Evaluation of Sixteen Algebraic Operations},
  author={Gomes, Renato Aparecido},
  doi={10.5281/zenodo.19325724},
  year={2026}
}

License

MIT