One‑bit binary codes can now match full‑precision retrieval quality, the authors report.
The authors surveyed approximate nearest‑neighbour techniques and reframed them under a three‑step lens: projection, quantisation and organisation. They released the open‑source BitBudget benchmark and ran reproducible tests on seven embedding models. A single‑bit code with full‑precision re‑ranking achieved the same recall as uncompressed vectors for six models, using just 1/32 of the memory. When byte budgets were equal, binary codes overtook a traditional inverted‑file product quantiser as the embedding size grew. Adding class labels to an eight‑byte supervised code more than doubled retrieval quality over a 2 KB task‑agnostic float.
This matters because retrieval‑augmented generation and large‑scale search have been fragmented across academic silos. Showing that aggressive quantisation can preserve, or even improve, quality challenges the assumption that higher‑dimensional floats are necessary for state‑of‑the‑art performance. It also gives practitioners a clear, low‑memory path to scale up services.
In short, the study suggests the community should consolidate around the projection‑quantisation‑organisation framework and consider binary codes as the default low‑memory choice for large‑scale retrieval.