Agent Eval

When to use

When the product has AI-powered features that produce non-deterministic output. Standard unit tests verify code logic; agent evals verify AI behavior.

Eval workflow

1. Define criteria: What does "correct" look like? Be specific.

   • "Returns relevant documents" → "Top 3 results include at least one document from the correct legal category"
   • "Good summary" → "Summary contains the case name, ruling, and at least 2 key arguments"

2. Create dataset: Build a set of (input, expected_output) pairs.

   eval_dataset = [
       {"input": "Find cases about land disputes in Nairobi", "expected": "must reference Land Act"},
       {"input": "What is the penalty for tax evasion?", "expected": "must cite Income Tax Act s.100"}
   ]
   

3. Run eval: Execute the agent on each input. Compare output to expected.

4. Score: Pass/fail per criterion. Overall pass rate.

5. Iterate: If pass rate < threshold, adjust prompts/tools/retrieval and re-eval.

What to eval in a RAG system

• Retrieval: Are the right documents retrieved? (precision + recall)
• Generation: Does the answer use the retrieved documents? (faithfulness)
• Relevance: Does the answer address the question? (answer relevance)

Anti-patterns

• "It seems to work" is NOT an eval
• Testing on 1-2 examples is NOT an eval (minimum 20 for meaningful signal)
• Evaluating only happy paths (must include edge cases, empty inputs, adversarial inputs)