Peichun Hua and co‑authors released a technique for private dense retrieval that uses a learned binary hash to pick a short candidate list. The shortlist is then reranked with encryption, hiding both the query and the final documents. On a 2.68 million‑passage dataset, the method adds just 0.38 seconds to a Qwen3‑32B RAG run over a 10‑Gbps link, matching full‑corpus quality with only 200‑500 candidates.
Why it matters
Enterprises can deploy private search over large document stores without noticeable slowdown.
