Projection and Quantisation: A Unifying View of Learning to Hash, from Random Projections to the RAG Era

The paper presents a unified framework for approximate nearest-neighbor search, focusing on the interplay of projection, quantization, and organization in retrieval methods. It introduces the BitBudget benchmark for reproducible measurement, highlighting that a one-bit code with full-precision re-ranking can match the quality of uncompressed methods while significantly reducing memory usage. The findings emphasize the effectiveness of learned embeddings and supervised codes in improving retrieval quality, which is crucial for practitioners optimizing performance in large-scale retrieval and retrieval-augmented generation systems.