Introduce Evaluator Experiment (#3133)

* feat: add exchanged messages in LLMCallCompletedEvent

* feat: add GoalAlignment metric for Agent evaluation

* feat: add SemanticQuality metric for Agent evaluation

* feat: add Tool Metrics for Agent evaluation

* feat: add Reasoning Metrics for Agent evaluation, still in progress

* feat: add AgentEvaluator class

This class will evaluate Agent' results and report to user

* fix: do not evaluate Agent by default

This is a experimental feature we still need refine it further

* test: add Agent eval tests

* fix: render all feedback per iteration

* style: resolve linter issues

* style: fix mypy issues

* fix: allow messages be empty on LLMCallCompletedEvent

* feat: add Experiment evaluation framework with baseline comparison

* fix: reset evaluator for each experiement iteraction

* fix: fix track of new test cases

* chore: split Experimental evaluation classes

* refactor: remove unused method

* refactor: isolate Console print in a dedicated class

* fix: make crew required to run an experiment

* fix: use time-aware to define experiment result

* test: add tests for Evaluator Experiment

* style: fix linter issues

* fix: encode string before hashing

* style: resolve linter issues

* feat: add experimental folder for beta features (#3141)

* test: move tests to experimental folder

src/crewai/experimental/evaluation/experiment/runner.py

@@ -0,0 +1,117 @@
+from collections import defaultdict
+from hashlib import md5
+from typing import Any
+
+from crewai import Crew
+from crewai.experimental.evaluation import AgentEvaluator, create_default_evaluator
+from crewai.experimental.evaluation.experiment.result_display import ExperimentResultsDisplay
+from crewai.experimental.evaluation.experiment.result import ExperimentResults, ExperimentResult
+from crewai.experimental.evaluation.evaluation_display import AgentAggregatedEvaluationResult
+
+class ExperimentRunner:
+    def __init__(self, dataset: list[dict[str, Any]]):
+        self.dataset = dataset or []
+        self.evaluator: AgentEvaluator | None = None
+        self.display = ExperimentResultsDisplay()
+
+    def run(self, crew: Crew, print_summary: bool = False) -> ExperimentResults:
+        self.evaluator = create_default_evaluator(crew=crew)
+
+        results = []
+
+        for test_case in self.dataset:
+            self.evaluator.reset_iterations_results()
+            result = self._run_test_case(test_case, crew)
+            results.append(result)
+
+        experiment_results = ExperimentResults(results)
+
+        if print_summary:
+            self.display.summary(experiment_results)
+
+        return experiment_results
+
+    def _run_test_case(self, test_case: dict[str, Any], crew: Crew) -> ExperimentResult:
+        inputs = test_case["inputs"]
+        expected_score = test_case["expected_score"]
+        identifier = test_case.get("identifier") or md5(str(test_case).encode(), usedforsecurity=False).hexdigest()
+
+        try:
+            self.display.console.print(f"[dim]Running crew with input: {str(inputs)[:50]}...[/dim]")
+            self.display.console.print("\n")
+            crew.kickoff(inputs=inputs)
+
+            assert self.evaluator is not None
+            agent_evaluations = self.evaluator.get_agent_evaluation()
+
+            actual_score = self._extract_scores(agent_evaluations)
+
+            passed = self._assert_scores(expected_score, actual_score)
+            return ExperimentResult(
+                identifier=identifier,
+                inputs=inputs,
+                score=actual_score,
+                expected_score=expected_score,
+                passed=passed,
+                agent_evaluations=agent_evaluations
+            )
+
+        except Exception as e:
+            self.display.console.print(f"[red]Error running test case: {str(e)}[/red]")
+            return ExperimentResult(
+                identifier=identifier,
+                inputs=inputs,
+                score=0,
+                expected_score=expected_score,
+                passed=False
+            )
+
+    def _extract_scores(self, agent_evaluations: dict[str, AgentAggregatedEvaluationResult]) -> float | dict[str,  float]:
+        all_scores: dict[str, list[float]] = defaultdict(list)
+        for evaluation in agent_evaluations.values():
+            for metric_name, score in evaluation.metrics.items():
+                if score.score is not None:
+                    all_scores[metric_name.value].append(score.score)
+
+        avg_scores = {m: sum(s)/len(s) for m, s in all_scores.items()}
+
+        if len(avg_scores) == 1:
+            return list(avg_scores.values())[0]
+
+        return avg_scores
+
+    def _assert_scores(self, expected: float | dict[str, float],
+                        actual: float | dict[str, float]) -> bool:
+        """
+        Compare expected and actual scores, and return whether the test case passed.
+
+        The rules for comparison are as follows:
+        - If both expected and actual scores are single numbers, the actual score must be >= expected.
+        - If expected is a single number and actual is a dict, compare against the average of actual values.
+        - If expected is a dict and actual is a single number, actual must be >= all expected values.
+        - If both are dicts, actual must have matching keys with values >= expected values.
+        """
+
+        if isinstance(expected, (int, float)) and isinstance(actual, (int, float)):
+            return actual >= expected
+
+        if isinstance(expected, dict) and isinstance(actual, (int, float)):
+            return all(actual >= exp_score for exp_score in expected.values())
+
+        if isinstance(expected, (int, float)) and isinstance(actual, dict):
+            if not actual:
+                return False
+            avg_score = sum(actual.values()) / len(actual)
+            return avg_score >= expected
+
+        if isinstance(expected, dict) and isinstance(actual, dict):
+            if not expected:
+                return True
+            matching_keys = set(expected.keys()) & set(actual.keys())
+            if not matching_keys:
+                return False
+
+            # All matching keys must have actual >= expected
+            return all(actual[key] >= expected[key] for key in matching_keys)
+
+        return False