Skip to main content
Check both the passages returned by search and the answers your persona gives. A successful upload or a relevant document appearing in search is only part of the check.

Before you test

  • Upload documents and wait for them to reach READY.
  • Connect a knowledge tool to the persona and confirm its selected folders.
  • Record the document revisions, folder IDs, and document IDs, along with the persona’s prompt, model, and tool configuration.
  • Choose questions with expected answers supported by your sources. Keep the other settings and question script constant when measuring a document change.

Inspect retrieved passages

Search a knowledge group directly to check the evidence available in that folder. Use the search knowledge group API from a terminal or your server, with ANAM_API_KEY and KNOWLEDGE_GROUP_ID set in your environment:
Replace the example question with one about your documents. The response is an array of matching passages. For each relevant result, check: For a document update, search for a distinctive changed fact and inspect the returned text. Check that superseded document IDs are no longer returned. For PDFs and tables, compare the passage with the original page for missing columns, units, or footnotes. Retrieve pricing, eligibility, and procedure sections separately. Check that each passage includes the restrictions needed to interpret it, even when the rest of the document is absent. If a rule is missing, keep it beside each affected fact or instruction and test again. A direct search checks one folder with your chosen query. The persona may choose a different query or search several folders through its tools, so also test the conversation itself.

Test persona answers

Start a fresh session after changing documents or configuration. Check that the persona uses the intended knowledge tool, then compare its answer with the approved source. Knowledge tool events can help confirm that a search ran or failed; an event reporting completion alone does not establish answer quality. Use this fictional test set as a starting point and adapt the expected behavior to your sources: Run the first three questions as a conversation to test how the persona handles changed context. Repeat important cases in fresh sessions and record failures as well as successes.

Diagnose a failed check

If the evidence used in the conversation is correct but the answer contradicts it, investigate answer generation and application rules. Decisions that must follow fixed rules may need validation in your application before an answer is delivered.

Record your results

Keep a test record outside the searchable knowledge folders. Distinguish a passage from a direct API test from evidence captured during the persona’s actual tool call. Copy this record for each test case: Record the settings and revisions with each test run. If measuring performance, track search duration separately from the time the persona takes to answer. One successful run does not prove reliability.

Next steps

Complete the upload checklist. Correct unresolved failures and repeat the affected tests before releasing the updated knowledge to users.
Last modified on September 10, 2026