Menu

Use case

Which knowledge sources actually help my support agent?

Mine your real support transcripts into an eval set, then measure which help-center pages, policies, and canned answers move real cases.

The problem

Support stacks accumulate knowledge, help-center pages, macros, policies , and nobody knows which of it actually resolves tickets. The closest proxy is usually a hunch.

The workflow

  1. Mine

    Turn production transcripts into an eval set. Weak labels are marked as derived, content-hash ids keep it stable, and PII warnings keep you honest.

  2. Baseline

    Measure the agent against the mined cases.

  3. Value

    Value each knowledge source against real cases with controls and confidence intervals.

  4. Select

    Keep what pays, reject what is redundant or harmful, and export the keepers to a knowledge-base manifest.

What it looks like

kno
# Mine real support transcripts into an eval set
kno mine --logs transcripts.jsonl --format jsonl-chat --mode immediate

# Then value candidate knowledge against real cases
kno value --evals mined.jsonl --pool help-center.jsonl \
  --baseline-run-id <run id>

The recipe

There is a full walkthrough in the cookbook: Value your Zendesk knowledge , the pattern works for any support stack that can export transcripts and help-center content.

Stages involved

Try it →