Embedding Compression via PCA Improves Domain Adaptation in Dense Retrieval

Jul 15, 2026

A new study demonstrates that applying Principal Component Analysis (PCA) to compress query embeddings can enhance retrieval performance in specialized domains. The approach was tested across 9 dense retrievers and 14 MTEB datasets, showing improved NDCG@10 scores in 75.4% of model-dataset pairs compared to standard embeddings. This method preserves domain-relevant features while removing non-discriminative components, providing a lightweight, training-free alternative to traditional domain adaptation methods.

Why it matters: This technique offers a simple and efficient way to improve dense retrieval in specialized domains without the need for costly annotation or retraining.

Full story at: arXiv Information Retrieval ↗