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

# Summaries

A summary looks across **many evaluation runs** for one agent and produces an AI-generated **diagnosis** of its strengths and weaknesses, a chart of how its metrics are trending, and a **prioritized list of recommendations** for improving it. Instead of reading run after run, you get the big picture and a clear list of what to fix first.

***

## Viewing a summary

Go to **Evaluations → Summaries** and select an **agent**. Every summary shows three things:

<Columns cols={2}>
  <Card title="Diagnosis" icon="stethoscope">
    An AI narrative of the agent's consistent strengths and recurring failure patterns.
  </Card>

  <Card title="Recommendations" icon="lightbulb">
    Specific, actionable prompt and skill suggestions. Each carries a **priority** chip (High / Med / Low) and a **metric tag**, and can be filtered by priority or metric.
  </Card>

  <Card title="Metrics over time" icon="chart-line">
    A run-by-run chart with the seven quality metrics as percentage lines plus Response Time in seconds, plotted by run date.
  </Card>
</Columns>

There are two ways to view a summary: an always-on **automatic** one, and an **on-demand** one for a date range.

***

## Automatic summary

With no date range selected, you see the agent's **automatic** summary. VIVI regenerates it **automatically after every evaluation run completes** — there's no button to press — analyzing the agent's **last 10 completed runs**. At the top you'll see "{N} runs analyzed", "Generated {time} ago", and a **stale badge** if it's more than 30 days old.

**While it's working**, the page shows:

* **"Generating summary…"** while VIVI is analyzing.
* **"Taking longer than expected"** after about five minutes.
* A message that a summary will be generated automatically after the next run, if generation failed or there isn't enough data yet.
* A **"Run your first evaluation"** call-to-action if the agent has no runs.

***

## On-demand summary (date range)

When you pick a **date range**, the page switches to an on-demand analysis of the runs in that window.

<Steps>
  <Step title="Pick a date range">
    Use a preset — **Last 7 days**, **Last 14 days**, or **Last 30 days** — or set a custom range. It analyzes the **last 10 completed runs** in that window.
  </Step>

  <Step title="Read the metrics immediately">
    The metrics for the selected period appear right away.
  </Step>

  <Step title="Generate the analysis">
    Click **Generate diagnosis & recommendations**. VIVI computes the diagnosis and recommendations in real time (this takes a few seconds).
  </Step>

  <Step title="Browse each run">
    After generating, use the **run selector** to view the diagnosis and recommendations for each individual run in the range.
  </Step>
</Steps>

<Note>
  On-demand analyses are temporary — they **expire after about 5 days**. You can regenerate them at any time, and your latest automatic summary is never affected.
</Note>

***

## Automatic vs. on-demand

<Columns cols={2}>
  <Card title="Automatic" icon="wand-magic-sparkles">
    Always reflects the latest state, regenerated after each run completes. Uses the agent's last 10 completed runs, with no date range.
  </Card>

  <Card title="On-demand" icon="calendar-days">
    Focuses on the last 10 completed runs in a date range you choose. Metrics appear immediately; the diagnosis and recommendations are generated only when you click the button.
  </Card>
</Columns>

***

## Best Practices

* Check the **automatic summary** after a batch of runs to see the current state of an agent.
* Sort recommendations by **priority** and tackle **High** items first — those move the most metrics.
* Use the **date-range** view to compare a "before" and "after" window around a prompt change.
* Pair the **Diagnosis** narrative with the **Metrics over time** chart: the chart shows the trend, the diagnosis explains *why*.
* If a summary says there wasn't enough data, run a couple more evaluations and come back — it updates on the next run.