How scores work
Quality scores are shown on a 0–100 scale, where higher is better. You’ll only ever see these values:| Displayed score | Color | Meaning |
|---|---|---|
| 100 | Green | Highest — excellent |
| 90 | Lime | Good, minor issues |
| 70 | Yellow | Mixed — noticeable gaps |
| 50 | Orange | Weak |
| 25 | Red | Poorest score |
| N/A | — | Not scored — skipped or couldn’t be assessed |
VIVI displays only the buckets 100, 90, 70, 50, or 25 for scored metrics — you won’t see in-between values like 80 or 63. An N/A is not a low score; open the metric detail to see whether it was skipped or errored.
The seven quality metrics
Groundedness
answer vs. evidence
Is every claim in the answer supported by the evidence the agent retrieved — from knowledge bases, tools, or attached files? Your main defense against hallucinations.
Skipped when there is no tool or retrieved evidence to check against.
Is every claim in the answer supported by the evidence the agent retrieved — from knowledge bases, tools, or attached files? Your main defense against hallucinations.
Skipped when there is no tool or retrieved evidence to check against.
Relevance
reply vs. question
Did each reply address what the user actually asked, staying on-topic without drifting into unrequested tangents?
Always scored.
Did each reply address what the user actually asked, staying on-topic without drifting into unrequested tangents?
Always scored.
Usefulness
answer vs. need
Is the answer actionable and complete enough for the user to move forward — versus vague, evasive, or padded?
Always scored.
Is the answer actionable and complete enough for the user to move forward — versus vague, evasive, or padded?
Always scored.
Correctness
answer vs. expected outcome
Are the agent’s factual claims accurate compared to the Expected Outcome you defined?
Skipped when no expected outcome is provided.
Are the agent’s factual claims accurate compared to the Expected Outcome you defined?
Skipped when no expected outcome is provided.
Instruction Following
behavior vs. system rules
Did the agent obey the rules in its prompt throughout the conversation?
Skipped when no system prompt is available.
Did the agent obey the rules in its prompt throughout the conversation?
Skipped when no system prompt is available.
Task Completion
goal achieved?
By the end of the conversation, was the user’s goal actually achieved end-to-end — not merely acknowledged or partly attempted?
Always scored.
By the end of the conversation, was the user’s goal actually achieved end-to-end — not merely acknowledged or partly attempted?
Always scored.
Tool Usage
tool calls vs. task
Did the agent choose the right tools, pass them good inputs, use what they returned, and recover from any tool errors?
Skipped when the conversation has no tool calls to assess.
Did the agent choose the right tools, pass them good inputs, use what they returned, and recover from any tool errors?
Skipped when the conversation has no tool calls to assess.
Response Time
Response Time
The eighth measure — but not a scored metric. Shown in seconds (not a percentage), it’s how long the agent took to answer. Agents configured to reason more deeply or run more steps may naturally be slower, so read Response Time alongside the quality metrics rather than on its own. Shows N/A when the value isn’t available.
Reasoning & Recommendations
For every scored metric you get two explanations:Reasoning
A short explanation of why the answer received its score, pointing to what worked or went wrong. If the metric was skipped, it says “Skipped: ” instead.
Recommendations
Specific, actionable suggestions for improving the agent — usually a prompt, knowledge base, or integration change. Shown whenever the score is below perfect; otherwise “No information available.”
Why a metric shows N/A
A metric displays N/A when it couldn’t be scored — either skipped or errored — and the other metrics are unaffected. Reasons you may see:- No evidence — Groundedness had no tool/retriever evidence to check against.
- No reference — Correctness had no expected outcome to compare to.
- No system prompt — Instruction Following had no prompt to check behavior against.
- Tool result unbound — a tool’s output wasn’t connected into the answer.
- Unparseable tool input — the input the agent sent to a tool couldn’t be understood.
- Evidence truncated — the evidence was too long and was cut off.
- Judge error or Timeout — the scorer couldn’t complete for this conversation.
Expect N/A on metrics you didn’t set up for — Correctness without an expected outcome, or Groundedness on a conversation with no tools. That’s normal, not a failure.
Best Practices
- Don’t chase a perfect 100 on every metric — focus on the ones that matter for your use case.
- Treat a low Groundedness score as a hallucination warning; check the agent’s knowledge and integrations.
- If Tool Usage is low, review the tool inputs and outputs in the conversation replay.
- Always read the Reasoning before acting — it tells you why a score is low.
- Apply high-priority Recommendations first, then re-run to confirm the improvement.