RAG Source Attestation
Bind each AI citation to a ZK proof of the exact docHash it claims to reference. Citation integrity holds across index rebuilds.
Three voices from the front line.
- IT / digital transformation
“We have no way to confirm a RAG answer's citation is really the latest, authentic version”
- AI engineering
“We can't detect an untrusted source slipping into the RAG pipeline”
- Audit / quality
“We need to reconstruct and explain where an AI answer came from at audit time”
Hand over the source, or just the facts?
Change what reaches the AI, and the leakage risk goes with it.
- query:
- How do I apply for parental leave?
- retrieved_doc:
- work-rules.pdf
- version:
- v2.1 (old)
- content:
- excerpt…
- source:
- shared drive
- agent:
- did:lemma:agent-rag-assistant
- modelId:
- claude-3.7-sonnet
- policyHash:
- 0x5c41…
- outputCommitment:
- work-rules @ v3.2 (latest)
- satisfiesPolicy:
- true
- sourceHash:
- 0x9b1d…
- lineageChain:
- [issue, index, cite]
- issuer:
- did:lemma:docs.internal
- ZK verified:
- ✓ VALID
Each source the AI cites is cryptographically bound to the exact document version that citation points to. A citation becomes a verifiable reference, not just a label. The original body reaches neither the index nor the answer; what leaves is only the fact that "this citation comes from the version valid at answer time." Even if the vector DB is rebuilt or the policy is revised, the citation proof attached to a past answer stays immutable.
See the technical details ↗Choose on three criteria.
Only work that needs all three at once — pass without exposing, independent verification, tamper-proof — is Lemma's domain.
| Method | Pass without exposing | Independent verification | Tamper-proof |
|---|---|---|---|
| Access control only | △ | ✗ | ✗ |
| Masking / anonymization | △ | ✗ | ✗ |
| Encryption only | ✓ | ✗ | ✗ |
| Lemma (ZK proof)the only one with all 3 | ✓ | ✓ | ✓ |
What's next
We enter through AI-adoption and citation-integrity support and a PoC, and stay alongside you through to operations.
- A 30-minute review — identify workflows where AI attaches citations to answers but you cannot later prove the source is authentic.
- Narrow to 1–2 decisions (results) to prove — e.g. "this citation came from the docHash of GL-2025-08, section C" — the citation facts bound to an answer. No answer logs or source documents required.
- Design connection and version-fixing — connection to your existing RAG / retrieval frameworks (LangChain, LlamaIndex, etc.), and version-fixing of the referenced document version.
- Prove one path via a (quote-based) PoC — confirm citation proofs work in one AI/RAG workflow.
- Hands-on support from rollout through operations — existing plan tiers (Civic / Critical / Compliance) serve only as a cost reference; the setup and pricing are designed together.
Tell us one AI workflow where "citations appear, but you can't prove they're real" applies, in the first 30 minutes. No answer logs or source documents required.
The bigger picture
The bigger picture this use case belongs to.
We map use scenarios across industries and workflows by the four axes.
See use scenarios for Verifiable AI in Solutions →TRY LEMMA
Run it yourself.
No sales call needed — start hands-on with Lemma's products.