zkDaily - 2026-04-13

零知识证明 zkDaily

ZKP Frontier Tracker 🎯

Mon

04.13

2026

zkRAG: Efficiently Proving RAG Retrieval in Zero Knowledge

Paper

https://eprint.iacr.org/2026/709

Xinyang Yang HNSW ANNS

Yang et al. proposed zkRAG in their paper, a ZK proof system for HNSW algorithm to verify faithful RAG retrieval execution with linear prover time.

Notes

RAG systems rely on vector retrieval, but server-side black-box operations can cause deviations, requiring consistency verification.
zkRAG designs a PIOP for HNSW algorithm to enable ZK proofs, ensuring retrieval faithfulness to committed database and index.
Prover time is linear in HNSW search trace length, achieving asymptotically optimal online efficiency.
Introduces new techniques like hybrid lookup argument and priority-queue update checker for improved efficiency.
Single-threaded proof for typical queries takes seconds, hundreds of times faster than baselines, with lightweight verification.
Security highlighted: protects server embeddings and index privacy, only outputs revealed, preventing arbitrary deviations.

Collected by @icerdesign

零知识证明 zkDaily

Q&A Deep Dive 💬

Mon

04.13

2026

Why does RAG retrieval need zero-knowledge proofs?

In RAG-as-a-Service, clients cannot see the server’s embedding database or HNSW index, so they cannot verify whether retrieval results are genuine. zkRAG provides service consistency proofs.

What does zkRAG mainly prove?

It proves that the returned results are consistent with running HNSW approximate nearest neighbor search on a fixed committed database, while preserving zero-knowledge.

Why is zkRAG efficient?

Its prover time is linear in the HNSW search trace length, achieving asymptotically optimal online prover efficiency while keeping verification succinct.

Prompted by @icerdesign