Design an agent that answers employee questions using internal documents. Explain ingestion, permission-aware retrieval, generation, citations, evaluation, cost and latency controls, and how the service scales to a large employee population. ### Constraints & Assumptions - Source permissions can change after a document is indexed. - Answers must be grounded in documents the requesting user may read. - The system should return a useful failure when evidence is insufficient. - Do not assume a particular agent framework or exact user count beyond the stated large-scale goal. ### Clarifying Questions to Ask - Which document systems, formats, and update frequencies are in scope? - Must answers quote exact passages, perform actions, or only retrieve information? - What latency, freshness, and confidentiality requirements apply to different document classes? ### What a Strong Answer Covers - Incremental ingestion, parsing, chunking, metadata, versioning, and deletion propagation. - Identity and access control enforced during retrieval, not only during ingestion. - Hybrid retrieval, reranking, context construction, grounded generation, and source citations. - A cost and latency budget using routing, caching, smaller models, parallel retrieval, and bounded context. - Offline and online evaluation, prompt-injection defenses, auditability, scaling, and graceful degradation. ### Follow-up Questions - How would you prevent a retrieved document from instructing the agent to reveal another document? - How would you reindex a changed embedding model without losing freshness? - What would you cache without leaking results across users?