Operations Process — Who Does What When
This document covers how to run the AI stack in your organization. AI is not 「install it and it runs itself」— it needs people to feed in material and judge the results to become useful.
The tutorial is 「get your first answer from one document」. This document is 「keep it useful over time」.
Roles
| Role | Account | Responsibilities |
|---|---|---|
| AI Operator | Admin console account (env) | Service status · document indexing · pipeline management |
| Field User | Platform account system | Ask the chatbot and judge answers |
| System Administrator | Server access | Installation · key rotation · backups |
The console account is one shared by the team → Change initial password.
Full workflow
Where arrows cross swimlanes, that's a handoff point.
현장 피드백 → 문서 보강 → 재색인 If it doesn't run, AI stalls at the level of documents you loaded initially.
Set up a channel to collect feedback — most loops break because there's nowhere
to send it.
1. Onboarding — Until it becomes useful
| Step | Task | Owner |
|---|---|---|
| 1 | Install → One-line install | System Administrator |
| 2 | Set up console account & keys → Change initial password | System Administrator |
| 3 | Check overall status green from home → Admin console | AI Operator |
| 4 | Choose which documents to load | AI Operator + Field |
| 5 | Index documents + validate with Query Test | AI Operator |
| 6 | Release to field | — |
Loading many documents at random doesn't help. Start with what the field actually asks about often — equipment manuals, work standards, troubleshooting history, in that order. Old documents teach the chatbot to answer wrong with confidence.
2. Operations — Regular tasks
| Frequency | Task | Owner |
|---|---|---|
| Daily | Home status summary (outages = 0?) | AI Operator |
| Weekly | Pipeline status · execution history | AI Operator |
| Weekly | GPU utilization · XID, LLM response time (P95) | AI Operator |
| Monthly | Re-index new document versions, collect wrong-answer cases | AI Operator + Field |
| Quarterly | Key rotation → Change password & API keys | System Administrator |
Everything is viewable in the admin console.
GPU hardware or driver error. Even if it's running now, the next symptom is usually node failure.
3. When an answer is wrong — Where to fix it
「The chatbot gave a bad answer」has multiple layers. Narrow it down in this order.
| Check | If true, then |
|---|---|
| 1. Is the source document indexed? | No → add the document (Documents) |
| 2. Does search find it? | No → try different search modes (hybrid/local/…) (Query Test) |
| 3. Search finds it but answer is wrong? | The source document is probably outdated — update it |
| 4. Answer is slow or cuts off? | LLM or GPU issue → check LLM and GPU screens |
Most complaints from the field that 「AI is dumb」are really missing or outdated documents.
Common failure points
| Symptom | Usual cause |
|---|---|
| 「We rolled it out but nobody uses it」 | Field questions don't match loaded documents |
| 「The analytics screen is empty」 | Feature pipeline is stuck — check pipeline, not the model |
| 「It confidently gives wrong answers」 | Old documents still in the index |
| 「It got slow」 | GPU contention or LLM queue — check P95 |
Related documents
- Tutorial — Make the chatbot answer from your documents
- AI Admin console — the screen you see every day
- Anomaly detection & incident response
- Database model — what gets stored where