Skip to content

How-to guide

Test and evaluate

Test your agent with real questions before making it available.

3 min

At a glance

Test your agent with real questions before making it available.

Before you start

  • Your agent is created and configured (tools + context). See Configure tools.

Steps

1. Open the test panel

From the agent editor, use the test panel (on the right side of the screen). Enter a question representative of your use case.

Example test questions:

  • "Who are our main suppliers?"
  • "Create a Product entity named Sensor X200."
  • "How many orders are currently in progress?"

2. Run the test

Click Send to run the test. The agent processes the question using the tools and context you configured.

Test success

3. Evaluate the response

Check the response quality on several criteria:

  • Relevance: does the response match the question asked?
  • Tools used: did the agent use the right tools (ontology search, documents, etc.)?
  • Tone and format: does the response follow the system instructions?
  • Accuracy: is the cited data correct?

4. Diagnose errors

In case of an error or unsatisfactory response, identify the cause using the execution details.

Test error

SymptomLikely causeSolution
Empty or off-topic responseOntology context too narrowBroaden the scope of accessible entities
Agent does not use the expected toolTool not enabledCheck the tool configuration
Execution errorConnection or permission issueCheck the workspace and API keys
Response too long or too vagueInsufficient system instructionsSpecify the format and boundaries in the prompt

5. Adjust and re-test

Fix the configuration (tools, context, system instructions) based on the results, then re-run the test. Repeat until you get satisfactory responses.

Tip
Test at least 3 different scenarios before considering the agent ready: a simple question, a complex question, and an out-of-scope question (to verify the agent correctly responds that it cannot help).

Expected outcome

What you get
Your agent is tested and validated on representative scenarios. You have confidence in the quality of its responses and it is ready for real-world use.

Limits and common issues

  • Response time depends on the question complexity and the number of enabled tools. An agent with fewer tools responds faster.
  • If the agent systematically returns errors, check that your workspace contains data (entities, documents) within the configured scope.
  • Configuration changes take effect immediately for the next test, without needing to reload the page.

Need help?

Write to us: Support and contact.

A reproducible evaluation set

Keep the question, workspace, saved configuration and expected answer. For Client–Commande:

CaseQuestionSuccess criterion
Known recordFind CLI-001Correct reference and verifiable evidence
Missing identifierFind CLI-ABSENTExplicit absence, no invention
AmbiguityOrders for ACMEClarification when customers share a name
Missing documentWhich delivery term applies?Limitation stated without inventing a clause
Action outside profileModify the orderNo writing with a reading profile

Compare tool arguments, evidence and final output for every run. Count correct answers, insufficiently supported answers and access errors separately. Technical session success does not establish business quality.

After changing the model, profile or sources, rerun these cases and compare with previous results. Avoid publishing based on a single successful example.

Did this page help you achieve your goal?