Cross-Encoder Reranking¶

Phentrieve uses a two-stage retrieval architecture that combines fast bi-encoder retrieval with precise cross-encoder reranking to improve ranking quality for HPO term matching.

How It Works¶

Stage 1: Dense Retrieval (Bi-Encoder)¶

The bi-encoder (BioLORD-2023-M) creates separate embeddings for queries and HPO terms, enabling fast similarity search across the entire ontology. This provides high recall but may occasionally rank similar-but-incorrect terms highly.

Stage 2: Protected Reranking (Cross-Encoder)¶

The cross-encoder (BAAI/bge-reranker-v2-m3) processes query-document pairs together, enabling more accurate relevance judgments. Phentrieve implements protected two-stage retrieval:

High-confidence matches are protected: Dense retrieval results with similarity ≥0.7 are preserved at top positions
Uncertain matches are refined: Lower-confidence candidates are reranked by the cross-encoder
Results are merged: Protected matches stay on top, reranked results fill below

This approach prevents cross-encoders from demoting correct cross-lingual matches due to lexical bias.

Default Model¶

BAAI/bge-reranker-v2-m3 (568M parameters)

Fine-tuned from BGE-M3 on multilingual datasets
Supports 100+ languages via XLM-RoBERTa architecture
Optimized for semantic matching without requiring translation

Reference: Chen et al. (2024). BGE M3-Embedding: Multi-Lingual, Multi-Functionality, Multi-Granularity Text Embeddings Through Self-Knowledge Distillation. arXiv:2402.03216.

Configuration¶

CLI Usage¶

# Enable reranking with default settings
phentrieve query --enable-reranker

# Specify a different cross-encoder model
phentrieve query --enable-reranker --reranker-model "ncbi/MedCPT-Cross-Encoder"

# Control number of candidates to rerank
phentrieve query --enable-reranker --rerank-count 50

Configuration File¶

# phentrieve.yaml
enable_reranker: true
reranker_model: "BAAI/bge-reranker-v2-m3"
rerank_candidate_count: 50
dense_trust_threshold: 0.7  # Protect matches above this similarity

Alternative Models¶

Other cross-encoder models compatible with Phentrieve:

Model	Domain	Notes
`ncbi/MedCPT-Cross-Encoder`	Biomedical	NCBI's medical text matcher
`cross-encoder/ms-marco-MiniLM-L-12-v2`	General	Fast, English-focused

GPU Acceleration

Cross-encoder reranking benefits significantly from GPU acceleration. Phentrieve automatically uses CUDA when available.