uhsr 0.1.3

Unified Hyperbolic Spectral Retrieval (UHSR) - a novel text retrieval algorithm combining lexical and semantic search.

---

Unified Hyperbolic Spectral Retrieval (UHSR)

Unified Hyperbolic Spectral Retrieval (UHSR) is a novel text retrieval algorithm that fuses lexical search (using BM25) with semantic search (using dense embeddings) into a unified, robust, and scalable system. It leverages advanced techniques such as logistic normalization, harmonic fusion, and spectral re-ranking based on graph Laplacian analysis to produce interpretable relevance scores within the [0,1] range.

Key Features

• Hybrid Retrieval: Combines BM25 for lexical scoring and dense vector semantic similarity for contextual understanding.
• Advanced Fusion: Uses logistic normalization and harmonic fusion to integrate multiple scoring signals.
• Spectral Re-Ranking: Employs spectral analysis (using the graph Laplacian and Fiedler vector) to boost central, highly relevant candidates.
• Metric Flexibility: Supports multiple semantic similarity metrics (cosine, euclidean, Mahalanobis) to suit various datasets.
• Interpretable Scores: Final relevance scores are normalized to the [0,1] range.
• Scalable: Designed to work with both small and large datasets using FAISS for fast approximate nearest neighbor search.

Overview

UHSR provides an end-to-end text retrieval pipeline that starts with raw documents and ends with a ranked list of documents. It first applies BM25 to perform fast lexical filtering, then computes semantic similarity using dense embeddings. The two scoring components are fused via a harmonic mean after logistic normalization, ensuring that both lexical and semantic aspects contribute effectively. Finally, a spectral re-ranking step based on graph Laplacian analysis refines the ranking by boosting documents that are centrally located among the top candidates.

Intended Use

UHSR is intended for research and educational purposes and can serve as a strong foundation for further development in text retrieval and natural language processing applications.

For more details, visit the GitHub repository.