County counsel stores ordinance embeddings in a vector database isolated from the public web tier. Why is that design significant for RAG security?

Storing ordinance embeddings in a vector database isolated from the public web tier matters because the vector store is the RAG knowledge backend; isolating it strengthens a security boundary between untrusted internet traffic and municipal knowledge. Attackers who can manipulate or exfiltrate the vector index can poison retrieval or steal sensitive ordinance-related content. Claiming isolated vector storage mitigates prompt injection by filtering user queries before retrieval fails because isolation is network separation of the knowledge backend, not input filtering — it does not inspect or sanitize user prompts, so malicious instructions that reach the retriever are still passed through. Saying separate embedding storage supports retrieval access control by encrypting ordinance text at rest fails because encryption at rest protects confidentiality of stored data, not per-user retrieval authorization — isolation governs network placement, while access control requires authentication and query-time permission filtering. Saying dedicated vector indices improve model output safety by blocking unapproved ordinance citations fails because index placement does not validate generation — blocking citations requires downstream output guardrails, while isolation's significance is the security boundary around the knowledge backend. Exam caveat: isolation alone is incomplete without authentication, least privilege, encryption, and monitoring of query and update paths into the vector store. Operational check: place the vector DB on a private network, allow only the RAG service identity to query, separate read and admin update roles, encrypt embeddings at rest, and alert on anomalous index modifications.