Analytics
What the Analytics tab measures, and how to use it to find what your model should learn next.
The Analytics tab shows how a deployed project is used: how many requests it gets, how fast it answers, which tools are used and which requests it couldn't answer. It is a Pro feature.
Choosing the period
Choose 24h, 7 days, 30 days or All time. Every number and list on the tab covers that period. The chart groups requests by hour, day or week to suit it.
The numbers
| Measure | Meaning |
|---|---|
| Requests | Requests the project's endpoints answered in the period. |
| Unique callers | Distinct callers, counted by a one-way hash of their network address. Studio never stores the address itself. Callers it can't identify aren't counted. |
| Latency | The model's average response time, with the median (p50) and the time 95% of requests came in under (p95). |
| Unanswered | The share of requests that got no tool call. |
Request volume charts requests over the period, and notes how many fell back to Gemini through the AI fallback.
The lists
| List | What it is for |
|---|---|
| Top tools | The tools called most often. A tool that never appears may be hard to reach, or not needed. |
| Top utterances | What people ask most. Make sure these are answered well. |
| Unanswered questions | Requests that matched no tool. These are the best candidates for your next training round. Curate for retraining opens them on the Retrain tab. |
| Browsers | The browsers and clients sending requests. |
| Countries | Where requests come from. Unknown covers requests with no location. |
Using it
A short weekly routine:
- Check Unanswered. A rising share means users are asking for things the model hasn't learned.
- Read Unanswered questions. Some will be things your tools can do, phrased in a way the model hasn't seen. Others will be requests for things your app doesn't do yet, which is useful product feedback.
- Choose Curate for retraining, correct the calls and retrain. See Retrain.
Requests you reject on the Retrain tab still count here. Analytics records what happened, whatever you later decide to train on.
Every Monday, Fernfly also emails a summary of the previous seven days to each account whose projects had requests: requests, unique callers, the AI fallback and unanswered rates, and the top tools. This email comes on every plan.
Revised