Skip to main content

Knowledge Base Analytics: Measuring Usage and Citation Patterns

Grounding only delivers value if teams can measure it. An active knowledge asset that never appears in runtime citations may be poorly bound, poorly scoped, or simply not useful. An asset that appears constantly in the wrong workflow may be too broad. Keeptrusts gives you enough evidence to answer those questions even without inventing a separate analytics layer.

Today, knowledge-base measurement comes from the evidence streams that already power governance: citation records, event data, history details, CLI output, and exports. Used together, those sources tell you which assets are being used, how often, and where grounded responses are actually happening.

Use this page when​

  • You want to measure whether active knowledge assets are being used in production.
  • You need a practical way to spot stale, over-broad, or underused assets.
  • You want to turn citation data into a routine improvement loop.

Primary audience​

  • Primary: Technical Engineers
  • Secondary: AI operations teams, technical leaders, governance reviewers

The problem​

Teams often stop at activation. They create an asset, promote it, bind it, and assume the rollout is done. That leaves a blind spot. If nobody checks runtime evidence, the asset may be active but irrelevant, or relevant but overshadowed by other context.

This becomes expensive in subtle ways. Token overhead increases because assets are available, but nobody knows whether they are improving answers. Important assets may be missing from citation trails even though the team thought they were central. Less useful assets may stay active for months because no one measures their actual usage.

Without analytics, knowledge governance becomes static. Teams maintain content, but they do not learn from how the content behaves in production.

The solution​

Keeptrusts gives you a distributed but usable measurement model.

  • kt kb citations shows usage for a specific asset or across assets over a time window
  • the Events stream exposes citation-bearing governance records for live traffic inspection
  • History details show how knowledge was used in session context
  • evidence exports let teams assemble review packets for audits and deeper analysis

That is enough to answer the practical questions that matter.

  • Which assets are cited most often?
  • Which assets are never cited despite being active?
  • Did usage change after a promotion?
  • Are the wrong agents or workflows relying on a given asset?
  • Are grounded workflows paying a token cost without seeing proportional value?

The important mindset shift is to treat citations as operational metrics, not only as audit artifacts. They tell you whether your knowledge design matches real behavior.

Implementation​

Start with one short recurring report that inspects citation volume and recent event patterns.

# Review recent citations across the environment
kt kb citations --since 24h --format json > kb-citations-24h.json

# Pull recent events that include citation data
kt events tail --last 50 --json | \
  jq 'select(.citations) | {event_id: .id, agent: .agent_id, citations: .citations}'

If you want a simpler per-asset check, run kt kb citations <asset-id> --since 7d for high-value assets after every major content change. Then compare the results before and after promotion. If the asset remains uncited, review its binding, active state, relevance, and whether a different asset is already satisfying that use case.

Use History when you need conversational context. Use Events when you need system-wide operational patterns. Use exports when you need to package evidence for reviews that extend beyond day-to-day debugging.

One useful analytic habit is to pair citation review with cost awareness. The architecture and spend docs make it clear that recalled knowledge adds input tokens. If an asset adds context weight but rarely appears in citation evidence, it may not deserve its runtime footprint.

Results and impact​

Knowledge analytics helps teams prune and improve their grounded corpus. Active assets stop being assumed-good and start being evaluated by actual use. That tends to produce a healthier library: fewer stale assets, clearer bindings, and stronger confidence that the most important documents are the ones shaping answers.

It also improves change management. After promotion, teams can verify that the new version is actually appearing in citations rather than only trusting that a command succeeded. Over time, this creates a feedback loop where knowledge is curated based on evidence instead of intuition alone.

Key takeaways​

  • Citation records are useful not just for audits, but for operational analytics.
  • Measure active assets by real usage, not by how important they looked during rollout.
  • Events, History, kt kb citations, and exports each provide a different analytic view.
  • Unused active assets are a maintenance and token-cost signal worth investigating.
  • Promotion should be followed by measurement, not treated as the end of the workflow.

Next steps​