AI Management Console
The AI stack comes with one management console. You monitor GPU load, LLM latency, and document indexing status in your browser without logging into the server.
This document is for first-time users. Work through the 10-minute walkthrough in order, then find screen-by-screen details as needed.
Field operators and system administrators. The console is mostly read-only, so accidental clicks rarely break anything. Only three places let you change data: document upload, document deletion, and feature backfill — all marked separately in this document.
Access and login
The address depends on your deployment. Confirm with your administrator which one applies.
| Deployment | Address |
|---|---|
| Domain-based installation | https://admin.<COMPANY_DOMAIN> |
| IP access on company network | http://<AI_SERVER_IP>:9080 |
Enter your user ID and password on the login screen. The default ID is admin; your administrator sets the password during installation.
The console has no per-user accounts. The entire operations team shares one account. Since there is no per-user audit trail, keep the password within your team.
You must log in again after 8 hours. Your screen URL is remembered, so you return to where you left off. Service restart may log everyone out — see Sessions keep dropping.
Menu and button text appears in English (only the home screen service cards have Korean descriptions). In the explanation below, actual English screen names are bold with Korean meaning alongside.
10-minute walkthrough
New users: follow this sequence. View only — make no changes.
Step 1 — Log in and view home
Logging in takes you straight to Home. A single-line status summary appears at the top.
정상 12 주의 0 중단 0 합계 12
This number is your entire AI stack health snapshot. If 중단 is 0, you can rest easy. Numbers update every 20 seconds automatically — no refresh needed.
Click the expand arrow on the right side of the summary line to see a list of individual service states. If you expand it, it stays expanded on your next visit.
Hover your mouse over that service in the list. Why it is yellow appears in a tooltip. For example, the service may be alive but its database connection down.
Step 2 — Scan the service cards
Below the summary line are service cards. One card = one service.
- Border color and status dot show whether it is healthy
- Click to enter that service's screen
- Some cards are not clickable — they are internal services you cannot reach directly from the browser. Think of them as status-only cards
Step 3 — Check GPU
Click GPU on the left sidebar. A table of GPU devices appears.
| Column | Meaning |
|---|---|
| Utilization | How busy the GPU is right now |
| Memory | How much the model is using |
| Temperature | High sustained temperature degrades performance |
| XID | Count of NVIDIA hardware or driver errors |
XID means GPU hardware or driver error. Left unattended, the typical next symptom is server down. If you see a non-zero number, note it exactly and report it to your system administrator.
Step 4 — Check LLM response speed
Click LLM on the sidebar. One row per model shows response speed and throughput.
When you hear "answers are slow lately," check the P95 column here (the 5th slowest among the last 100 requests). If it is much higher than usual, the complaint is valid. The small line chart at the end shows recent trend.
Step 5 — Enter one service
Click RAG API on the sidebar. Its menu unfolds on the left. Click Documents. A list of indexed documents appears.
Do not click anything here — just note how many documents are indexed. When the chatbot says "I could not find related documents," this is the first place to check.
Step 6 — View logs
In the same service menu, click Logs. Recent logs appear in newest-first order (latest at the top). You do not need to SSH into the server and run docker logs.
Step 7 — Done
You have now seen 90% of what the console does. In summary:
Check overall status on home → navigate to the suspicious service → read the logs.
Find other screens in the reference below as needed.
Screen map
Left sidebar (always visible)
| Menu | Function |
|---|---|
| Home | Overall status summary + service cards |
| GPU | GPU device utilization, memory, temperature, XID |
| LLM | Model response speed and throughput |
| RAG API | Document search engine — indexing, sync, query testing |
| Timeseries Insight | Equipment anomaly detection and forecasting |
| Document Parsing | Service to extract tables and fields from documents |
| Vision | Image reading service (defect classification, gauge reading) |
| Operations Manual ↗ | This manual you are reading (new window) |
GPU and LLM appear above the services intentionally. Most issues trace down from infrastructure, so scanning top-to-bottom finds problems faster.
Operations Manual points to an external site (docs.plantpulse.io). It won't open where internet is unavailable. A local copy of the manual is bundled with the server installation — contact your administrator.
Four services share identical menus
RAG API · Timeseries Insight · Document Parsing · Vision all have the same four menus. Learn once, apply everywhere.
| Menu | Function |
|---|---|
| MCP Server Info | List of tools this service offers to the AI copilot |
| API Endpoints | List of APIs this service provides (for integration developers) |
| Logs | Recent logs (newest first) |
| Settings | Current configuration values — read-only |
It only displays values; it cannot change them. Passwords and API keys appear masked. To actually change settings, see Credentials and Operations.
RAG API — Document search
This service lets the chatbot answer based on company documents. When you hear "the chatbot can't find documents," start here.
| Menu | Function |
|---|---|
| Documents | List of indexed documents. Upload and delete are available here |
| Sync History | When sync ran and whether it succeeded |
| Query Test | Enter search terms to test directly what matches |
| Common 4 menus | See table above |
When chatbot cannot find a document
- Check Documents to see if the document is in the list — if not, indexing failed
- If it is there, use Query Test with language similar to what the user asked
- If still no match, try different search modes — you can choose among
hybrid,local,global,naive,mix. Default ishybrid - Check Sync History to verify the last sync did not fail
The chatbot stops using a deleted document as evidence. To use it again, you must re-upload the original file. Do not delete if you do not have the original file.
Timeseries Insight — Anomaly detection and forecasting
This engine finds anomaly signals in equipment data and predicts the future. It has the most menus.
Analysis screens
Each category pairs Anomaly Detection with Forecast.
| Category | Screens |
|---|---|
| Asset (equipment) | Anomaly Detection · Forecast · Health Index |
| Production | Anomaly Detection · Quality Forecast |
| Alarm | Frequency Anomaly · Failure Forecast · Duration Analysis · Sensor Correlation · Statistics |
| System | Anomaly Detection · Forecast |
| Profiling | Asset Profiling — deep dive into one equipment |
Feature Store — the fuel for analysis
Analyses above run on processed data stacked in the feature store. Empty or stale results usually mean the problem is here.
| Menu | Function |
|---|---|
| Pipeline Status | Is data processing running normally now |
| Run History | Past execution history and pass/fail |
| Feature Data | Query the actual stored values |
| Backfill | Recalculate a missing period |
It recalculates historical periods, so wide ranges take time and raise server load. Run it during low-production hours, with the narrowest range needed. For first-time use, consult your administrator.
Document Parsing · Vision
Both services have only the four common menus. In the console, you check they are alive and review logs when issues arise.
- Document Parsing — extracts layout, tables, and fields from documents
- Vision — classifies defects from images or reads gauges
Both are internal services unreachable directly from the browser, so their home screen cards are not clickable. The console polls their status on your behalf.
Troubleshooting
| Symptom | Look first | Then |
|---|---|---|
| Chatbot can't find documents | RAG API → Documents | Query Test for direct search → Sync History |
| Chatbot answers are slow | LLM P95 column | GPU utilization and temperature |
| Analysis results are empty | Insight → Pipeline Status | Run History failed rows |
| Yellow light on home | Expand status list, check tooltip | That service → Logs |
| Red light on home | That service → Logs | Operations restart procedure |
| Screen is blank with error banner | Service may not be up yet | Check status on home |
The console shows an error banner, not a blank table, when data cannot be fetched. An error banner means connection failed, not missing data.
Accounts and sessions
Console accounts are managed via environment variables, not a database.
| Variable | Meaning | Default |
|---|---|---|
ADMIN_AUTH_USERNAME | Login user ID | admin |
ADMIN_AUTH_PASSWORD | Login password | None — must be set |
ADMIN_SESSION_SECRET | Session signing key | If empty, generated fresh on each restart |
ADMIN_SESSION_MAX_AGE | Login session duration (minutes) | 480 (8 hours) |
If ADMIN_AUTH_PASSWORD is empty, the console refuses to start. This prevents an unauthenticated admin portal from quietly running — normal behavior.
If ADMIN_SESSION_SECRET is empty, each service restart generates a new signing key, logging everyone out. Set a value and logins survive restarts.
bin/passwd.shConsole credentials are not in the key management list that bin/passwd.sh handles — see the table in Credentials. To change it, edit the environment file directly and restart the service. Contact your administrator.
Related documents
- Service port map — console address and where each service listens
- Operations — startup, shutdown, restart procedures
- Credentials — key rotation
- Logs and monitoring — reading logs directly on the server