> ## Documentation Index
> Fetch the complete documentation index at: https://docs.chatropic.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Conversation Evals

> Score real conversations against quality criteria

Open **Evals** under **Agent settings** (`/settings/evals`). Conversation Evals score your agent's **real logged conversations** against a fixed set of quality criteria, so you can see not just whether conversations were resolved, but how well they were handled, and which part of the system to fix when they were not.

## Run an evaluation

<Steps>
  <Step title="Pick a date range">
    Choose **Last 7 days**, **Last 30 days**, or a custom range. The page shows how many logged conversations in that range are eligible for scoring (a conversation needs at least one user turn and one agent turn).
  </Step>

  <Step title="Run evals">
    Click **Run evals**. Each eligible conversation is scored against every criterion, and the summary cards show the **Overall score**, the weakest criterion, and the run status.
  </Step>

  <Step title="Review low scores">
    Open individual conversations from the results table to see the per-criterion verdicts and the specific issues detected.
  </Step>
</Steps>

## The criteria

Each criterion maps to a part of the system, which tells you where to make the fix:

| Criterion                   | Owner      | Low score usually means                                                         |
| --------------------------- | ---------- | ------------------------------------------------------------------------------- |
| **Retrieval quality**       | Knowledge  | The right document chunks were not found; add or restructure sources            |
| **Answer accuracy**         | Generation | Replies did not match your content; fix or split the underlying documents       |
| **Conversation continuity** | Memory     | The agent lost the thread or restarted mid-conversation                         |
| **Lead capture execution**  | Conversion | The agent missed chances to collect contact details it was configured to gather |
| **Redundancy**              | Prompt     | Replies repeat phrasing or content enough to feel mechanical                    |
| **Conciseness**             | Prompt     | Answers run longer than the conversation needs                                  |

## Make it a habit

Evals are most useful as a routine, not a one-off:

* Run them **after any significant change**: new knowledge sources, a new or edited [Skill](/user-guides/agent-management/skills), or changed guardrails.
* Compare the overall score across runs. A drop after a change points straight at that change.
* Pair with the [Topic Explorer](/user-guides/analytics/topic-explorer): topics with negative sentiment and low eval scores are the same problem seen from two angles.

<Note>
  Conversation Evals score conversations that already happened. For pre-launch testing before real traffic exists, use **Playground** with a **New test session** and run through your expected journeys manually. Per-skill behavior can also be tested with [skill simulations](/user-guides/agent-management/skills#test-with-simulations).
</Note>

<CardGroup cols={2}>
  <Card title="AI performance" icon="chart-line" href="/user-guides/analytics/ai-performance">
    Volume, resolutions, and CSAT trends.
  </Card>

  <Card title="Best practices" icon="star" href="/user-guides/quick-start/best-practices">
    Fixes for the most common low-score causes.
  </Card>
</CardGroup>