White paper WP-022
RAG Citation Integrity: Per-Chunk ZK Proofs for Agentic AI
Per-chunk retrieval proofs that let auditors verify every citation without trusting your vector database.
At a glance
- Paper
- WP-022
- Topic
- RAG Citation Integrity: Per-Chunk ZK Proofs for Agentic AI
- Format
- PDF + web summary
- Signatures
- ML-DSA-65 post-quantum (NIST FIPS 204)
- Sandbox
- Reproducible at affix-io.com/sandbox
- Company
- AffixIO, Wales, UK
RAG Citation Integrity: Per-Chunk ZK Proofs for Agentic AI is an AffixIO technical paper. Per-chunk retrieval proofs that let auditors verify every citation without trusting your vector database.
Agentic RAG systems cite sources they never retrieved. Regulators notice. This paper walks through AffixIO's per-chunk zero-knowledge retrieval proofs: each cited passage is cryptographically bound to a Merkle-anchored document chunk before the model sees it. Auditors verify citations offline. Article 13 transparency stops being a slide deck promise.
Summary
Agentic RAG systems cite sources they never retrieved. Regulators notice. This paper walks through AffixIO's per-chunk zero-knowledge retrieval proofs: each cited passage is cryptographically bound to a Merkle-anchored document chunk before the model sees it. Auditors verify citations offline. Article 13 transparency stops being a slide deck promise.
Download the full PDF for technical detail, diagrams, and reproduction steps. Public sandbox: affix-io.com/sandbox.
Related reading
Frequently asked questions
What is RAG citation integrity?
It means every citation in an agentic RAG response can be traced to a specific retrieved chunk, with cryptographic proof that chunk existed in your corpus at query time.
How do zero-knowledge proofs help RAG?
They let you prove a chunk was retrieved and hashed without exposing the full document store or user queries to third-party auditors.
Does this satisfy EU AI Act requirements?
Per-chunk attestation supports Article 13 transparency and Article 12 record-keeping by producing verifiable retrieval evidence, not just model output logs.