53 lines
1.5 KiB
Markdown
53 lines
1.5 KiB
Markdown
|
# Request Analysis On Debian 12 VPS
|
||
|
|
|
||
|
|
This setup runs the request-analysis worker on the VPS only when needed.
|
||
|
|
|
||
|
|
## How It Works
|
||
|
|
|
||
|
|
- the API server accepts pending launcher requests
|
||
|
|
- when a new pending request is created, the server schedules the worker
|
||
|
|
- only one worker run is active at a time
|
||
|
|
- if more pending requests remain after a run finishes, the server schedules another pass
|
||
|
|
- you can also trigger a run manually through `POST /api/launcher-requests/process-pending`
|
||
|
|
|
||
|
|
The worker itself still uses DeepSeek's hosted API, so the VPS is coordinating queue work instead of doing local inference.
|
||
|
|
|
||
|
|
## Required Environment
|
||
|
|
|
||
|
|
Recommended env file:
|
||
|
|
|
||
|
|
`/srv/worldshaper/shared/worldshaper.env`
|
||
|
|
|
||
|
|
Example:
|
||
|
|
|
||
|
|
```bash
|
||
|
|
DEEPSEEK_API_KEY=your-key-here
|
||
|
|
REQUEST_ANALYZER_AUTORUN=1
|
||
|
|
REQUEST_ANALYZER_PROVIDER=deepseek
|
||
|
|
REQUEST_ANALYZER_MODEL=deepseek-v4-flash
|
||
|
|
REQUEST_ANALYZER_PROMOTE_THRESHOLD=0.85
|
||
|
|
REQUEST_ANALYZER_THINKING=disabled
|
||
|
|
```
|
||
|
|
|
||
|
|
## PM2 Deploy Hook
|
||
|
|
|
||
|
|
The sample post-receive hook in [scripts/vps-post-receive.sample.sh](../scripts/vps-post-receive.sample.sh) now:
|
||
|
|
|
||
|
|
- loads `/srv/worldshaper/shared/worldshaper.env` if present
|
||
|
|
- restarts PM2 with `--update-env`
|
||
|
|
- passes the request-analysis environment through on restart/start
|
||
|
|
|
||
|
|
## Manual Trigger
|
||
|
|
|
||
|
|
If you want to force a queue run:
|
||
|
|
|
||
|
|
```bash
|
||
|
|
curl -X POST http://127.0.0.1:5180/api/launcher-requests/process-pending
|
||
|
|
```
|
||
|
|
|
||
|
|
## Notes
|
||
|
|
|
||
|
|
- autorun is opt-in through `REQUEST_ANALYZER_AUTORUN=1`
|
||
|
|
- the worker only launches if pending requests exist
|
||
|
|
- DeepSeek credentials stay on the VPS in the shared env file, not in the repo
|