Skip to main content

Observability

Evaluations

Testing an agent before it answers a real caller.

What evaluation means on Finn

Finn does not have an automated evaluation suite. There is no scored test-set runner, no regression harness, and no API endpoint that grades an agent. Evaluation today means placing test calls yourself, reading the transcript and the extracted fields, and fixing what you find before the agent takes a real caller. A programmatic evaluation surface is not yet settled.

That makes the loop manual, so the discipline matters more. The sections below are the order to work in.

Before the first test call

Check these in the dashboard. Each one fails in a way that looks like a model problem but is not.

CheckWhereFailure if skipped
Welcome message has text, or is set to dynamicEdit Finn → Welcome MessageAgent picks up and says nothing
Voice assigned, region matchesEdit Finn → VoiceSilence or wrong-sounding output
Knowledge base shows IndexedKnowledge Base tabAgent answers "I don't know" to documented facts
Post-call analysis fields definedPost-call analysis configFields stay blank for every call in this test round
Trial credit remainingSettings → Plan and billing → UsageTest call never dials

Post-call analysis is the one that bites hardest. Fields only populate on calls that complete after the field was added. If you add a field mid-test, every call you already placed stays blank forever, and no re-run backfills them. Define your fields first, then test. See post-call-analysis.

Placing the test call

Enter the number with a country code. US is +15551234567, India is +919876543210. The platform shows a "tested format: …" preview before it triggers, so read that preview rather than trusting the number you typed.

A test call that never arrives is usually not the agent:

CauseCheck
Number on the global suppression listSettings → Compliance → Do Not Call List
Carrier rejectionSome international carriers block new numbers. Try a different from-number.
Spam filter on your handsetApple and Google filters drop calls silently. Try another from-number.
From-number caller ID not registered (US)Settings → Phone numbers → [Number] → Registration Status

Full symptom list in troubleshooting.

Reading the result

Listen to the recording. Do not reason from the transcript alone, and do not guess from the summary. Find the first turn that went wrong, not the worst one. Everything after the first bad turn is downstream noise.

Then open the Data Extractor Sidebar for that call:

TabWhat it tells you
KnowledgeWhich document snippets were retrieved. Wrong snippets mean a knowledge base problem, not a prompt problem.
Extraction StatusWhether post-call fields ran and whether extraction errored.

If the retrieved snippets were correct and the agent still answered wrong, the bug is in the prompt. If the snippets were wrong or empty, fix the knowledge base first. See agents-knowledge-base.

Fixing and re-testing

The loop from the prompting guide applies directly:

  1. Find the first weird turn.
  2. Add a rule covering that exact scenario, plus one example turn.
  3. Re-test on the same number.
  4. If it is still wrong, the cause is upstream in the workflow, the audience, or the knowledge base.

Most "the model is dumb" reports are missing-rule bugs. Details in agents-prompting.

What a test call cannot tell you

Be honest about the limits, because a clean test call is weak evidence.

SignalWhy it misleads
SentimentAn indicator, not a verdict. It misclassifies. If it is consistently wrong for your use case, add a custom post-call field with a specific question and use that instead.
Very short callsCalls under five seconds are treated as no-answer by some metrics and will not appear where you expect. Check the full call log.
One clean passYou tested the happy path. Test the angry caller, the ambiguous request, and the caller who says "remove me" and expects the call to end.
Your own reactionsYou know the script. A real caller does not, and will interrupt, mumble, and change topic.

Test the negative rules explicitly. If the prompt says never quote a price, ask for a price. If it says never book a Sunday, ask for Sunday. A rule that has never been provoked has never been verified.

After you go live

Testing does not stop at launch. Watch the first real calls in call-logs and the deployment health panel in monitoring. The failure modes that only appear at volume, such as hangups from an unregistered caller ID or an exhausted audience, will not show up in a single test call.