Controlled ingestion
Authenticate sources, scan/validate content, capture source/version metadata and reject unsupported or unexpectedly large inputs.
Open reference architecture
Secure RAG Architecture
A retrieval architecture that preserves source permissions, provenance and evaluation instead of treating a vector database as an unrestricted knowledge dump.
Permission-aware indexing
Carry access rules into retrieval; a vector index must not become a route around source-system permissions.
Chunk provenance
Retain source identifiers and enough lineage to trace an answer back to the material retrieved.
Retrieval boundaries
Filter by tenant, user entitlement, sensitivity and purpose before content reaches the model.
Prompt-injection resistance
Treat retrieved documents as data. Separate system policy from retrieved text and prevent documents from granting themselves tool authority.
Answer grounding
Prefer citations/evidence, expose uncertainty and abstain when retrieval does not support a material claim.
Evaluation
Test retrieval quality, permission leakage, injection resistance, unsupported claims and changes after model/index updates.
Retention & deletion
Define how source deletion propagates to chunks, caches, embeddings and generated artefacts where required.