Securing a RAG Pipeline

Try it out loud first. Then check yourself:

1. Enforce document ACLs at retrieval time using the requesting user's identity; the vector DB must filter by permission.
2. Ingestion is an attack surface: poisoned or injection-laden docs become trusted context. Vet and label sources.
3. Embeddings are not anonymized: inversion attacks can reconstruct text, so protect the vector store like the source data.
4. Multi-tenant stores need hard tenant isolation (separate namespaces/indexes), not just metadata filters in prompts.
5. Log retrieved chunk IDs per answer for auditability, and apply output filters for PII and secrets.

If the interviewer pushes back

• A document's permissions change in SharePoint. How quickly does your RAG index reflect it, and what is the risk window?
• Maps to which OWASP LLM Top 10 item, and what does that list miss for RAG?

Go deeper

• owasp.org/www-project-top-10-for-large-language-model-applications/