The Top questions tab shows what callers ask in the calls a cohort scores. Similar questions are grouped, each group gets one representative question, and you see how often it was asked and how often the calls that asked it were resolved. Use it to find gaps in your prompt or knowledge base: a common question with a low resolution rate is the first thing to fix.
What the tab shows#
Three tiles:
| Tile | Meaning |
|---|---|
| Calls analyzed | Scored calls in the cohort |
| Calls with questions | Scored calls where at least one question was found |
| Pass rate | Share of scored calls that pass, after overrides |
Then a table of question groups, most asked first:
| Column | Meaning |
|---|---|
| Question | The group's label, with one example of how a caller phrased it |
| Asked | How many times questions in this group were asked across all calls |
| Calls | How many calls asked at least one question in this group |
| Resolved | Share of those calls that were resolved (after overrides) |
The tab shows up to 20 groups. "No questions yet" means no scored call has a question.
How questions are found#
When a call is scored, Tried reads the caller's turns and splits them into sentences. A sentence counts as a question when, after dropping leading filler such as "um", "so", "okay" or "hi", it is at least 6 characters long and either ends with ? or starts with a question word (what, when, where, why, who, how, which, can, could, do, does, is, are, will, would, should, may, have, has). Each call keeps up to 20 distinct questions, stored on the QA result as questions.
Only the caller's words are used. Questions the agent asks are not counted.
How questions are grouped#
Questions are grouped by keyword overlap. Each question's significant words (longer than 2 letters, without common words) are compared with every existing group; it joins the group with the highest overlap if that overlap is at least 0.3 (Jaccard similarity), otherwise it starts a new group. A group's keywords grow as questions join it.
The label is the most frequent question in the group (the shortest wins a tie).
Needs keys With an OpenAI key, gpt-5.6-terra writes one short canonical question for each group from its examples. Without a key, or if that request fails, the most frequent question is used.
Reading the results#
- High Asked, low Resolved: callers keep asking something the agent cannot answer. Add the answer to the knowledge base or the prompt, or route these calls to a person.
- High Asked, high Resolved: a common, well-handled request. Consider making it faster, for example by answering it in the welcome message.
- Questions you did not expect: check whether the agent's scope or the number's routing is right.
Resolution here is the QA verdict, not post-call call_successful. Overrides you make on the QA sheet change it.
API#
GET /qa/cohorts/:id/top-questions?limit=20 (limit 1 to 50, default 20):
{
"calls_analyzed": 148,
"calls_with_questions": 97,
"pass_rate": 0.716,
"questions": [
{
"question": "Do you have parking?",
"count": 31,
"calls": 29,
"resolution_rate": 0.517,
"examples": ["Do you have parking?", "Is there parking near the clinic?", "Where can I park?"]
}
]
}examples holds up to 5 distinct phrasings. resolution_rate and pass_rate are fractions from 0 to 1.