Pointing the Way, Hiding the Destination: Practical Private Dense Retrieval at Scale
Ranking
Overall
91
Content
100
Popularity
71
Observed public metrics from 1 member.
Merged summary
TL;DR - This paper presents a privacy-preserving dense retrieval protocol for hosted RAG and semantic search that uses learned deep hashing to privately narrow searches before encrypted reranking. It approaches full-corpus retrieval quality while adding only 0.73 seconds to a Qwen3-32B RAG pipeline on a 2.68M-passage corpus.
- Randomized binary codes produce short candidate lists while hiding queries and reducing full-corpus cryptographic computation.
- Encrypted reranking and oblivious key transfer protect the exact query, selected result, and access authorization.
- Shortlists of 200–500 candidates closely match full-corpus retrieval across five zero-shot datasets ranging from 25K to 5.4M documents.
- The released implementation satisfies directional metric differential privacy and reduces embedding-inversion and property-inference leakage.
Sources (1)
Pointing the Way, Hiding the Destination: Practical Private Dense Retrieval at Scale
Public signals
Semantic Scholar citations 1 · Semantic Scholar influential citations 0
TL;DR - This paper presents a privacy-preserving dense retrieval protocol for hosted RAG and semantic search that uses learned deep hashing to privately narrow searches before encrypted reranking. It approaches full-corpus retrieval quality while adding only 0.73 seconds to a Qwen3-32B RAG pipeline on a 2.68M-passage corpus.
- Randomized binary codes produce short candidate lists while hiding queries and reducing full-corpus cryptographic computation.
- Encrypted reranking and oblivious key transfer protect the exact query, selected result, and access authorization.
- Shortlists of 200–500 candidates closely match full-corpus retrieval across five zero-shot datasets ranging from 25K to 5.4M documents.
- The released implementation satisfies directional metric differential privacy and reduces embedding-inversion and property-inference leakage.