Deep explanation
Secure RAG Quality Improvement in a customer onboarding assistant
Improve RAG quality in an onboarding assistant while staying secure: document-level access control in retrieval, indirect prompt injection from uploaded and web content, tool authorization, PII handling, and joint quality and security evaluation.
Quick answer
Every new source adds both useful context and a new attack surface, so I would classify each source by trust level and sensitivity before indexing it. Access control has to happen inside the retrieval query, with customer and tenant filters, never by asking the model to ignore other customers' data. Untrusted text such as uploads and web pages is treated as data that can never trigger tools, and tools are authorized by backend code. Then I would measure retrieval quality and security together, so a quality gain that leaks data or follows injected instructions does not ship.
The interview question
To improve RAG answer quality in a customer onboarding assistant, your team wants to add more sources: customer-uploaded documents, CRM notes, help-center articles, and partner web pages. Security review flagged prompt injection, data leakage between customers, and unauthorized tool access. How would you improve retrieval quality without opening these security holes?
Deep explanation
Simple explanation
Sign in to unlock the full answer
Free accounts include 5 full deep answers. Sign in to start unlocking.
- 22 more sections of deep explanation
- Real-world examples
- Common mistakes
- Interviewer expectations
- Follow-up questions