Skip to content

Commit 80f062b

Browse files
authored
Merge pull request #64 from tao12345666333/doc-reasoning-model-ctl
docs: add reasoning model metrics
2 parents 8a1f76b + 4089a3b commit 80f062b

File tree

1 file changed

+43
-0
lines changed

1 file changed

+43
-0
lines changed

website/docs/api/router.md

Lines changed: 43 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -175,6 +175,49 @@ semantic_router_pii_detections_total{action="block"} 23
175175
semantic_router_jailbreak_attempts_total{action="block"} 5
176176
```
177177

178+
### Reasoning Mode Metrics
179+
180+
The router exposes dedicated Prometheus counters to monitor reasoning mode decisions and template usage across model families. These metrics are emitted by the router and can be scraped by your Prometheus server.
181+
182+
- `llm_reasoning_decisions_total{category, model, enabled, effort}`
183+
- Description: Count of reasoning decisions made per category and selected model, with whether reasoning was enabled and the applied effort level.
184+
- Labels:
185+
- category: category name determined during routing
186+
- model: final selected model for the request
187+
- enabled: "true" or "false" depending on the decision
188+
- effort: effort level used when enabled (e.g., low|medium|high)
189+
190+
- `llm_reasoning_template_usage_total{family, param}`
191+
- Description: Count of times a model-family-specific template parameter was applied to requests.
192+
- Labels:
193+
- family: normalized model family (e.g., qwen3, deepseek, gpt-oss, gpt)
194+
- param: name of the template knob applied (e.g., enable_thinking, thinking, reasoning_effort)
195+
196+
- `llm_reasoning_effort_usage_total{family, effort}`
197+
- Description: Count of times a reasoning effort level was set for a given model family.
198+
- Labels:
199+
- family: normalized model family
200+
- effort: effort level (e.g., low|medium|high)
201+
202+
Example PromQL:
203+
204+
```prometheus
205+
# Reasoning decisions by category and model (last 5m)
206+
sum by (category, model, enabled, effort) (
207+
rate(llm_reasoning_decisions_total[5m])
208+
)
209+
210+
# Template usage by model family and parameter (last 5m)
211+
sum by (family, param) (
212+
rate(llm_reasoning_template_usage_total[5m])
213+
)
214+
215+
# Effort distribution by model family (last 5m)
216+
sum by (family, effort) (
217+
rate(llm_reasoning_effort_usage_total[5m])
218+
)
219+
```
220+
178221
## gRPC ExtProc API
179222

180223
For direct integration with the ExtProc protocol:

0 commit comments

Comments
 (0)