LLM endpoints
Configure an OpenAI-compatible endpoint. A cron worker feeds each conversation to it to detect whether a deal was made, which unlocks peer reviews.
How many conversations are checked in parallel each run, load-balanced across active endpoints (each endpoint capped by its Concurrent value).
| label | URL | Model ID | Concurrent | Used for | Active |
|---|
News AI: priority, queue and statistics
On endpoints used for both deal checks and news, news leaves this many of the endpoint’s concurrent slots for deal checks (news still gets at least one), and starts no news step while deal checks are waiting. 0 = no reservation.
News AI queue
| Title | Step | Status | Endpoint | Ready in |
|---|
Statistics
| Endpoint | Step | Done | Failed | Average | Slowest | Speed |
|---|
Speed is generated tokens per second (reported by the endpoint). "Ready in" estimates use these averages.
Add endpoint
Leave “system role” off for models like Gemma that reject it (instructions are folded into the message). Turn on “disable thinking” for hybrid reasoning models like Qwen3 so they answer directly.
Choose at least one. News rewriting writes long articles: give the model server a large context (16K tokens or more).