Summary

Monitor the age of content across an agent's knowledge sources at the space level, with drill-down to individual pages. Alert administrators when content becomes stale, optionally exclude stale content from retrieval, and display a freshness indicator to users on agent responses.

Problem Statement

Rovo agents are only as good as the content they draw from. When knowledge source pages are not updated, the agent continues to serve outdated information with the same confidence as current information. Users have no way to distinguish a response based on a recently updated page from one based on a page last touched 18 months ago.

For enterprise environments with agents indexing hundreds or thousands of pages across multiple spaces, this creates a compounding risk:

1. Silent accuracy drift. Content staleness degrades accuracy without triggering any alert or re-evaluation.

1. User trust erosion. When users discover the agent gave them outdated information, they lose trust in all of the agent's responses.

1. No accountability signal. Requesting teams are responsible for maintaining their knowledge sources, but there is no mechanism to alert them when content is aging or to show the impact of unmaintained content on agent quality.

1. Unmanageable at page level. An agent indexing thousands of pages cannot be monitored page-by-page. Freshness must be reported at the knowledge source (space) level with drill-down available when needed.

Proposed Solution

Knowledge freshness page (administrator view)

A new "Knowledge freshness" item in the Agent Studio sidebar.

Freshness thresholds (configurable per agent):

Threshold Default Meaning Aging 180 days Content is approaching staleness. Yellow indicator shown to users. Stale 365 days Content is considered outdated. Red indicator shown to users. Optionally excluded from retrieval.

Administrators can adjust thresholds based on the agent's domain. An agent covering a rapidly changing topic might use 90/180 day thresholds, while an agent covering stable reference material might use 365/730 days.

Enforcement and notification toggles:

• Exclude stale content from retrieval (default: off) - Pages exceeding the stale threshold are removed from the agent's retrieval pool. The agent behaves as if the content does not exist. If no other source covers the topic, the agent deflects rather than serving outdated information. Pages are automatically re-included when updated.

• Show freshness rating to users (default: off) - Display a freshness indicator on agent responses when the source content exceeds the aging threshold.

• Notify knowledge source owner on stale content (default: on) - Send a notification to the requesting team's nominated owner when pages in their knowledge source exceed the stale threshold.

Space-level overview

The primary view is a table showing each knowledge source (Confluence space, Jira space, or external source) with:

• Space name and key

• Total page count

• Freshness distribution - A stacked bar showing the percentage of pages that are fresh (green), aging (yellow), stale (red), and excluded (grey)

• Stale page count - Number of pages exceeding the stale threshold

• Excluded page count - Number of pages currently excluded from retrieval

• View pages - Drill-down link to see individual pages

This gives administrators a manageable view regardless of how many pages the agent indexes. The conversation with a content owner is "your space has 18% aging content and 8% excluded" rather than "here are 124 individual pages."

Space drill-down

Clicking "View pages" on a space opens a filtered page-level table showing:

• Page title (linked to the Confluence page)

• Last updated date

• Age in days

• Status - Fresh, Aging, Stale, or Excluded

• Reference count - How many times the agent retrieved this page in the last 30 days

The drill-down defaults to showing stale and aging pages only (sorted by age, oldest first), since these are the actionable items. Filters allow switching to all pages, excluded only, or fresh only. Sort options include age, reference count, and page title.

Excluded pages show a struck-through reference count, visually communicating that this page is no longer being served and indicating the volume of queries that are now going unanswered on that topic.

Content exclusion

When the "Exclude stale content from retrieval" toggle is enabled, pages exceeding the stale threshold are automatically removed from the agent's retrieval pool. The thresholds apply uniformly - there is no per-page configuration needed.

This is a deliberate trade-off: the agent may deflect on topics where it previously had an answer, but a deflection ("I wasn't able to find current information on this topic") is a better outcome than confidently serving information from a page that hasn't been reviewed in over a year.

When a page is updated and returns within the freshness threshold, it is automatically re-included in the retrieval pool without administrator intervention.

Freshness indicator (user view)

When the "Show freshness rating to users" toggle is enabled, agent responses display a small badge below the response indicating the age of the source content:

• Fresh (green badge): "Source updated X days ago"

• Aging (yellow badge): "Source updated X days ago"

• Stale (red badge): "Source last updated X days ago - this information may be outdated"

The badge includes a link to the source page, allowing the user to verify the information directly.

When content has been excluded, the agent deflects naturally without a freshness badge: "I wasn't able to find current information on this topic. The available documentation may be out of date and has been temporarily excluded. Please contact the platform support team for the latest process."

Design Reference

See attached mockup (mockup-knowledge-freshness.html) showing:

1. Mockup 1 - Space-level overview with freshness settings (thresholds, exclusion toggle, notification toggle), and the knowledge source table showing four spaces with stacked freshness distribution bars, stale/excluded counts, and drill-down links.

1. Mockup 2 - Space drill-down into ATLCLOUD showing stale and aging pages with filters, sort options, age, status tags, and reference counts. Excluded pages shown with struck-through reference counts.

1. Mockup 3 - User view showing three chat interactions: a fresh response (green badge), an aging response (yellow badge), and a deflection where the source was excluded (no badge, natural deflection message).

Use Cases

Space-level content health check: The platform team opens the knowledge freshness page and sees that the ATLCLOUD space has 18% aging and 8% excluded content. They notify the requesting team's content owner with a single data point: "your space needs review - 10 pages are currently excluded from the agent."

Enforcing content maintenance via exclusion: Content exclusion is enabled with a 365-day stale threshold. The Account Creation Guide, referenced 47 times per month, exceeds the threshold and is excluded. Users who ask about account creation now receive a deflection. The requesting team, alerted by both the notification and the increase in "Didn't answer my question" user feedback, prioritises updating the page. Once updated, the page is automatically re-included.

User self-service verification: A user asks about group permissions. The response includes a yellow aging badge indicating the source was last updated 224 days ago. The user clicks the source link, checks the content, and determines whether the guidance is still current.

Evaluation correlation: A monthly evaluation shows accuracy has dropped from 90% to 82%. The platform team checks knowledge freshness and finds that three frequently referenced pages have crossed the stale threshold since the last evaluation. The accuracy drop is explained by content staleness, not an agent configuration issue.

Interaction with Other Features

• User feedback: "Out of date" feedback correlates directly with knowledge source staleness. The platform team can cross-reference feedback reports with the freshness data.

• Evaluation datasets: When content is updated after a staleness alert, the requesting team should review their accuracy dataset to ensure expected responses still match the updated content.

• Compliance dashboard: Consider adding a knowledge freshness indicator to the compliance dashboard, showing the percentage of fresh/aging/stale/excluded content per agent.

Considerations

• Freshness is not correctness. A page last updated 2 years ago may still be perfectly accurate (e.g. a reference architecture document). Staleness is a risk indicator, not a definitive quality signal. Threshold adjustment accommodates this - an agent covering stable content can use longer thresholds.

• Exclusion impact. When content exclusion is enabled, administrators should review the space-level stale and excluded counts to understand the impact before enabling. The stacked bar visualisation makes this immediately visible.

• Reference count dependency. The reference count column requires Atlassian to track which pages the agent retrieves and cites in responses. If this data is not currently available, the feature can operate without it, but reference counts significantly increase the feature's value by highlighting high-impact staleness.

• Threshold inheritance. Consider allowing site-level default thresholds (via the Governance tab in Rovo settings) that apply to all agents, with per-agent overrides.

• Space-level notification. Notifications should aggregate at the space level rather than firing per-page. A single notification saying "ATLCLOUD has 3 new pages that crossed the stale threshold this month" is actionable. 124 individual page notifications is noise.