Making the AI assistant guide a non-technical admin (lightweight RAG, not a vector DB)
The operator found the UI hard and wanted the AI to guide them. Added a doc-grounded help_docs tool so the assistant answers how-to/install questions strictly from a curated usage guide — lightweight RAG, no vector DB needed at this scale.
The person running DocVault is a non-technical admin. The web UI was "too hard," and the natural fix wasn't to redraw every screen — it was to let them just ask the AI "how do I install this on an employee's PC?" and get a correct, step-by-step answer.
The gap
The assistant was a data assistant: tool-use over live rows (event_counts, list_users, list_alerts, …) plus action tools (create user, assign host, ack alert). It had no usage knowledge — ask it a how-to and it had nothing to ground on, so it would flounder or risk inventing steps. Usage docs existed (manual.ko.html) but the agent never read them.
"Do we need RAG?"
The operator's instinct was right: answers must be fact-based, from our docs, not freestyled. That's exactly what RAG is — retrieve, then generate. But "RAG" doesn't have to mean a vector database + embeddings. That machinery earns its keep when the corpus is too big to fit in context. Our usage guide is a couple of KB.
So the accurate, simple choice: a curated guide (internal/agent/help_kb.md, embedded) and a help_docs tool that returns it. The model reads the whole guide and answers grounded in it. For a small KB this is more accurate than vector retrieval — there's no retrieval step to miss, and the LLM itself handles paraphrase ("how do I install it" ↔ the guide's "에이전트 설치"). We upgrade to ranked/embedding retrieval only when the KB stops fitting in context.
What shipped (commit c3c2c69)
help_kb.md: a curated, plain-Korean how-to guide written against the actual UI (login, register users, CSV import, install on employee PCs incl. the SmartScreen steps, host assignment, viewing risk, acknowledging alerts, using the assistant).help_docstool wired intoNewEngine; system prompt now says: for usage/install/how-to questions, callhelp_docsfirst and answer step-by-step in simple Korean grounded only in that text; if it's not in the guide, say so — don't invent.- Dashboard AI box now invites how-to questions ("직원 컴퓨터에 어떻게 깔아?").
The existing injection defense still holds: the help guide is trusted content the developer wrote; tool output from monitored PCs remains untrusted and can't trigger actions.
Lesson
"We need RAG" usually means "ground it in our facts." Do that — but match the retrieval to the corpus. For a small, curated KB, returning the guide is the retrieval; a vector DB would be over-engineering.