User Feedback Loop
Wire production traffic back into your model selection logic. This recipe shows how to capture thumbs-up / thumbs-down signals from end users, persist them against the Meridian request id, and use the aggregate to bias the router toward models that actually delight your customers.
1.Capture the signal at the edge
Every Meridian completion returns an x-meridian-request-id header. Echo it into your UI, then POST it back to the feedback endpoint when the user reacts. The id is the join key between the model that served the response and the rating it earned.
- Surface a thumbs-up / thumbs-down control on every assistant turn.
- Capture the request id from the response headers, never from client state.
- Fire-and-forget the rating — do not block the next user message on it.
2.Persist with the request id
POST to /v1/feedback with the request id, a polarity, and an optional free-text comment. Meridian joins this against the original routing decision so the analytics dashboard can show per-model satisfaction over any time window.
curl -X POST https://llm.getnimbus.net/v1/feedback \
-H "Authorization: Bearer $MERIDIAN_KEY" \
-H "Content-Type: application/json" \
-d '{
"request_id": "req_8f3a2c1e",
"polarity": "positive",
"comment": "Picked the right tone for support reply",
"tags": ["support", "tone"]
}'3.Close the loop in routing
Once you have a week of signal, flip router.feedback_bias on in your project settings. Meridian will down-weight models whose 7-day thumbs-down rate exceeds your threshold and quietly promote the ones your users prefer — without you touching prompts or pricing.
Tip: Start with a soft bias (10–15%) so a few bad ratings don't evict a model entirely. Tune upward once you trust the volume.