A clear view of what HardRAG processes and records.
HardRAG is designed for regulated AI review workflows. This page explains the current data model, deployment options, audit approach, and limitations in plain language.
What HardRAG processes
Validation requires the information needed to judge a RAG answer: the prompt, retrieved context, generated answer, applicable policies, and metadata.
What HardRAG stores
Cloud and sovereign deployment paths.
Cloud mode
Intended for fast pilots and managed deployments using Supabase-backed authentication, database, storage, and dashboard workflows.
Sovereign mode
Designed for local or air-gapped environments with local LLMs and customer-controlled infrastructure. Some components are roadmap-dependent and should be scoped per deployment.
Tamper-evident today. Stronger ledger controls on the roadmap.
The current implementation records an integrity hash with evaluation metadata. Production hardening should add hash-chain sequencing, signing, verification endpoints, and append-only storage controls.
SHA-256 style integrity hash on audit result fields.
Structured audit record with scores, rationales, privacy data, and judge output.
Hash-chain or ledger-style sequencing for stronger tamper evidence.
Signed exports and verification endpoint for audit packages.