Evaluating & Testing Matches
Use the evaluation tool to check how inputs are matched against your training data.
Before you publish, it's worth checking that your training data actually catches the answers people will give. The Evaluate training data tool lets you type a sample answer and see exactly how the matching engine responds â so you can fix gaps before real participants hit them.
Run an evaluation
Open the tool
In the Train module choose Evaluate, and select the question block you want to test.
Enter a sample answer
Type an answer a participant might give in the Question field, then click Start evaluation.
Read the results
The matching report shows which records were ranked, with a clear Best match at the top. Matches that scored too low appear under Highest match but below cutoff â these would not trigger during a real dialog.
Reading the matching report
What you see | What it means |
|---|---|
Best match | The record the bot would use for this answer. Its Reaction will be used. |
Highest match but below cutoff | The closest record, but not close enough to be used â a sign you may need more Input variants. |
Global training data | Indicates the match came from your global records rather than the question block. |
If a sensible answer lands below the cutoff, add it as an Input variant of the right record, then evaluate again.
Test the LLM and chat behaviour
Two extra buttons go further than plain matching:
Test LLM response â generates an AI response based on your existing training data, useful for open questions and quizzes.
Simulate chat behaviour â plays out how the bot would actually behave in the chat, including any Action such as Move to the next question or Stay on the current question.
LLM generation needs example training data to work from. If you only have records for correct answers but test an incorrect one, no LLM response can be generated â so cover both cases.
When the matches look right, move on to testing the full dialog.