kailash 0.3.2__py3-none-any.whl → 0.4.1__py3-none-any.whl

kailash/nodes/rag/evaluation.py

@@ -0,0 +1,875 @@
+"""
+RAG Evaluation and Benchmarking Framework
+
+Implements comprehensive evaluation metrics and benchmarking:
+- Retrieval quality metrics (precision, recall, MRR)
+- Generation quality assessment
+- End-to-end RAG evaluation
+- A/B testing framework
+- Performance benchmarking
+- Dataset generation for testing
+
+Based on RAGAS, BEIR, and evaluation research from 2024.
+"""
+
+import json
+import logging
+import random
+import statistics
+import time
+from datetime import datetime
+from typing import Any, Dict, List, Optional, Tuple, Union
+
+from ...workflow.builder import WorkflowBuilder
+from ..ai.llm_agent import LLMAgentNode
+from ..base import Node, NodeParameter, register_node
+from ..code.python import PythonCodeNode
+from ..logic.workflow import WorkflowNode
+
+logger = logging.getLogger(__name__)
+
+
+@register_node()
+class RAGEvaluationNode(WorkflowNode):
+    """
+    Comprehensive RAG Evaluation Framework
+
+    Evaluates RAG systems across multiple dimensions including retrieval
+    quality, generation accuracy, and end-to-end performance.
+
+    When to use:
+    - Best for: System optimization, quality assurance, model selection
+    - Not ideal for: Real-time evaluation during inference
+    - Performance: 5-30 seconds per evaluation (depends on metrics)
+    - Insights: Detailed breakdown of strengths and weaknesses
+
+    Key features:
+    - RAGAS-based evaluation metrics
+    - Retrieval and generation quality assessment
+    - Faithfulness and relevance scoring
+    - Comparative analysis across strategies
+    - Automated test dataset generation
+
+    Example:
+        evaluator = RAGEvaluationNode(
+            metrics=["faithfulness", "relevance", "context_precision", "answer_quality"],
+            use_reference_answers=True
+        )
+
+        # Evaluate a RAG system
+        results = await evaluator.run(
+            test_queries=[
+                {"query": "What is transformer architecture?",
+                 "reference": "Transformers use self-attention..."},
+                {"query": "Explain BERT",
+                 "reference": "BERT is a bidirectional..."}
+            ],
+            rag_system=my_rag_node
+        )
+
+        # Results include:
+        # - Per-query scores
+        # - Aggregate metrics
+        # - Failure analysis
+        # - Improvement recommendations
+
+    Parameters:
+        metrics: List of evaluation metrics to compute
+        use_reference_answers: Whether to use ground truth
+        llm_judge_model: Model for LLM-based evaluation
+        confidence_threshold: Minimum acceptable score
+
+    Returns:
+        scores: Detailed scores per metric
+        aggregate_metrics: Overall system performance
+        failure_analysis: Queries that performed poorly
+        recommendations: Suggested improvements
+    """
+
+    def __init__(
+        self,
+        name: str = "rag_evaluation",
+        metrics: List[str] = None,
+        use_reference_answers: bool = True,
+        llm_judge_model: str = "gpt-4",
+    ):
+        self.metrics = metrics or [
+            "faithfulness",
+            "relevance",
+            "context_precision",
+            "answer_quality",
+        ]
+        self.use_reference_answers = use_reference_answers
+        self.llm_judge_model = llm_judge_model
+        super().__init__(name, self._create_workflow())
+
+    def _create_workflow(self) -> WorkflowNode:
+        """Create RAG evaluation workflow"""
+        builder = WorkflowBuilder()
+
+        # Test executor - runs RAG on test queries
+        test_executor_id = builder.add_node(
+            "PythonCodeNode",
+            node_id="test_executor",
+            config={
+                "code": """
+import time
+from datetime import datetime
+
+def execute_rag_tests(test_queries, rag_system):
+    '''Execute RAG system on test queries'''
+    test_results = []
+
+    for i, test_case in enumerate(test_queries):
+        query = test_case.get("query", "")
+        reference = test_case.get("reference", "")
+
+        # Time the execution
+        start_time = time.time()
+
+        # Execute RAG (simplified - would call actual system)
+        # In production, would use rag_system.run(query=query)
+        rag_response = {
+            "answer": f"Generated answer for: {query}",
+            "retrieved_contexts": [
+                {"content": "Context 1 about transformers...", "score": 0.9},
+                {"content": "Context 2 about attention...", "score": 0.85},
+                {"content": "Context 3 about architecture...", "score": 0.8}
+            ],
+            "confidence": 0.87
+        }
+
+        execution_time = time.time() - start_time
+
+        test_results.append({
+            "test_id": i,
+            "query": query,
+            "reference_answer": reference,
+            "generated_answer": rag_response["answer"],
+            "retrieved_contexts": rag_response["retrieved_contexts"],
+            "execution_time": execution_time,
+            "timestamp": datetime.now().isoformat()
+        })
+
+    result = {
+        "test_results": test_results,
+        "total_tests": len(test_queries),
+        "avg_execution_time": sum(r["execution_time"] for r in test_results) / len(test_results)
+    }
+"""
+            },
+        )
+
+        # Faithfulness evaluator
+        faithfulness_evaluator_id = builder.add_node(
+            "LLMAgentNode",
+            node_id="faithfulness_evaluator",
+            config={
+                "system_prompt": """Evaluate the faithfulness of the generated answer to the retrieved contexts.
+
+Faithfulness measures whether the answer is grounded in the retrieved information.
+
+For each statement in the answer:
+1. Check if it's supported by the contexts
+2. Identify any hallucinations
+3. Rate overall faithfulness
+
+Return JSON:
+{
+    "faithfulness_score": 0.0-1.0,
+    "supported_statements": ["list of supported claims"],
+    "unsupported_statements": ["list of unsupported claims"],
+    "hallucinations": ["list of hallucinated information"],
+    "reasoning": "explanation"
+}""",
+                "model": self.llm_judge_model,
+            },
+        )
+
+        # Relevance evaluator
+        relevance_evaluator_id = builder.add_node(
+            "LLMAgentNode",
+            node_id="relevance_evaluator",
+            config={
+                "system_prompt": """Evaluate the relevance of the answer to the query.
+
+Consider:
+1. Does the answer address the query?
+2. Is it complete?
+3. Is it focused without irrelevant information?
+
+Return JSON:
+{
+    "relevance_score": 0.0-1.0,
+    "addresses_query": true/false,
+    "completeness": 0.0-1.0,
+    "focus": 0.0-1.0,
+    "missing_aspects": ["list of missing elements"],
+    "irrelevant_content": ["list of irrelevant parts"]
+}""",
+                "model": self.llm_judge_model,
+            },
+        )
+
+        # Context precision evaluator
+        context_evaluator_id = builder.add_node(
+            "PythonCodeNode",
+            node_id="context_evaluator",
+            config={
+                "code": """
+def evaluate_context_precision(test_result):
+    '''Evaluate the precision of retrieved contexts'''
+
+    contexts = test_result.get("retrieved_contexts", [])
+    query = test_result.get("query", "")
+
+    if not contexts:
+        return {
+            "context_precision": 0.0,
+            "context_recall": 0.0,
+            "context_ranking_quality": 0.0
+        }
+
+    # Calculate precision at different k values
+    precision_at_k = {}
+    relevant_count = 0
+
+    for k in [1, 3, 5, 10]:
+        if k <= len(contexts):
+            # Simulate relevance judgment (would use LLM in production)
+            relevant_at_k = sum(1 for c in contexts[:k] if c.get("score", 0) > 0.7)
+            precision_at_k[f"P@{k}"] = relevant_at_k / k
+
+    # Calculate MRR (Mean Reciprocal Rank)
+    first_relevant_rank = None
+    for i, ctx in enumerate(contexts):
+        if ctx.get("score", 0) > 0.7:
+            first_relevant_rank = i + 1
+            break
+
+    mrr = 1.0 / first_relevant_rank if first_relevant_rank else 0.0
+
+    # Context diversity
+    unique_terms = set()
+    for ctx in contexts:
+        unique_terms.update(ctx.get("content", "").lower().split()[:20])
+
+    diversity_score = len(unique_terms) / (len(contexts) * 20) if contexts else 0
+
+    result = {
+        "context_metrics": {
+            "precision_at_k": precision_at_k,
+            "mrr": mrr,
+            "diversity_score": diversity_score,
+            "avg_relevance_score": sum(c.get("score", 0) for c in contexts) / len(contexts),
+            "context_count": len(contexts)
+        }
+    }
+"""
+            },
+        )
+
+        # Answer quality evaluator (if reference available)
+        if self.use_reference_answers:
+            answer_quality_id = builder.add_node(
+                "LLMAgentNode",
+                node_id="answer_quality_evaluator",
+                config={
+                    "system_prompt": """Compare the generated answer with the reference answer.
+
+Evaluate:
+1. Factual accuracy
+2. Completeness
+3. Clarity and coherence
+4. Additional valuable information
+
+Return JSON:
+{
+    "accuracy_score": 0.0-1.0,
+    "completeness_score": 0.0-1.0,
+    "clarity_score": 0.0-1.0,
+    "additional_value": 0.0-1.0,
+    "overall_quality": 0.0-1.0,
+    "key_differences": ["list of major differences"],
+    "improvements_needed": ["list of improvements"]
+}""",
+                    "model": self.llm_judge_model,
+                },
+            )
+
+        # Metric aggregator
+        aggregator_id = builder.add_node(
+            "PythonCodeNode",
+            node_id="metric_aggregator",
+            config={
+                "code": f"""
+import statistics
+
+def aggregate_evaluation_metrics(test_results, faithfulness_scores, relevance_scores,
+                               context_metrics, answer_quality_scores=None):
+    '''Aggregate all evaluation metrics'''
+
+    # Parse evaluation results
+    all_metrics = {{
+        "faithfulness": [],
+        "relevance": [],
+        "context_precision": [],
+        "answer_quality": [],
+        "execution_time": []
+    }}
+
+    for i, test in enumerate(test_results):
+        # Get scores for this test
+        faith_score = faithfulness_scores[i].get("response", {{}}).get("faithfulness_score", 0)
+        rel_score = relevance_scores[i].get("response", {{}}).get("relevance_score", 0)
+        ctx_score = context_metrics[i].get("context_metrics", {{}}).get("avg_relevance_score", 0)
+
+        all_metrics["faithfulness"].append(faith_score)
+        all_metrics["relevance"].append(rel_score)
+        all_metrics["context_precision"].append(ctx_score)
+        all_metrics["execution_time"].append(test.get("execution_time", 0))
+
+        if answer_quality_scores:
+            quality_score = answer_quality_scores[i].get("response", {{}}).get("overall_quality", 0)
+            all_metrics["answer_quality"].append(quality_score)
+
+    # Calculate aggregate statistics
+    aggregate_stats = {{}}
+    for metric, scores in all_metrics.items():
+        if scores:
+            aggregate_stats[metric] = {{
+                "mean": statistics.mean(scores),
+                "median": statistics.median(scores),
+                "std_dev": statistics.stdev(scores) if len(scores) > 1 else 0,
+                "min": min(scores),
+                "max": max(scores),
+                "scores": scores
+            }}
+
+    # Identify failure cases
+    failure_threshold = 0.6
+    failures = []
+
+    for i, test in enumerate(test_results):
+        overall_score = (all_metrics["faithfulness"][i] +
+                        all_metrics["relevance"][i] +
+                        all_metrics["context_precision"][i]) / 3
+
+        if overall_score < failure_threshold:
+            failures.append({{
+                "test_id": i,
+                "query": test["query"],
+                "overall_score": overall_score,
+                "weakest_metric": min(
+                    ("faithfulness", all_metrics["faithfulness"][i]),
+                    ("relevance", all_metrics["relevance"][i]),
+                    ("context_precision", all_metrics["context_precision"][i]),
+                    key=lambda x: x[1]
+                )[0]
+            }})
+
+    # Generate recommendations
+    recommendations = []
+
+    if aggregate_stats.get("faithfulness", {{}}).get("mean", 1) < 0.7:
+        recommendations.append("Improve grounding: Ensure answers strictly follow retrieved content")
+
+    if aggregate_stats.get("relevance", {{}}).get("mean", 1) < 0.7:
+        recommendations.append("Enhance relevance: Better query understanding and targeted responses")
+
+    if aggregate_stats.get("context_precision", {{}}).get("mean", 1) < 0.7:
+        recommendations.append("Optimize retrieval: Improve document ranking and selection")
+
+    if aggregate_stats.get("execution_time", {{}}).get("mean", 0) > 2.0:
+        recommendations.append("Reduce latency: Consider caching or parallel processing")
+
+    result = {{
+        "evaluation_summary": {{
+            "aggregate_metrics": aggregate_stats,
+            "overall_score": statistics.mean([
+                aggregate_stats.get("faithfulness", {{}}).get("mean", 0),
+                aggregate_stats.get("relevance", {{}}).get("mean", 0),
+                aggregate_stats.get("context_precision", {{}}).get("mean", 0)
+            ]),
+            "failure_analysis": {{
+                "failure_count": len(failures),
+                "failure_rate": len(failures) / len(test_results),
+                "failed_queries": failures
+            }},
+            "recommendations": recommendations,
+            "evaluation_config": {{
+                "metrics_used": {self.metrics},
+                "total_tests": len(test_results),
+                "timestamp": datetime.now().isoformat()
+            }}
+        }}
+    }}
+"""
+            },
+        )
+
+        # Connect workflow
+        builder.add_connection(
+            test_executor_id, "test_results", faithfulness_evaluator_id, "test_data"
+        )
+        builder.add_connection(
+            test_executor_id, "test_results", relevance_evaluator_id, "test_data"
+        )
+        builder.add_connection(
+            test_executor_id, "test_results", context_evaluator_id, "test_data"
+        )
+
+        if self.use_reference_answers:
+            builder.add_connection(
+                test_executor_id, "test_results", answer_quality_id, "test_data"
+            )
+            builder.add_connection(
+                answer_quality_id, "response", aggregator_id, "answer_quality_scores"
+            )
+
+        builder.add_connection(
+            test_executor_id, "test_results", aggregator_id, "test_results"
+        )
+        builder.add_connection(
+            faithfulness_evaluator_id, "response", aggregator_id, "faithfulness_scores"
+        )
+        builder.add_connection(
+            relevance_evaluator_id, "response", aggregator_id, "relevance_scores"
+        )
+        builder.add_connection(
+            context_evaluator_id, "context_metrics", aggregator_id, "context_metrics"
+        )
+
+        return builder.build(name="rag_evaluation_workflow")
+
+
+@register_node()
+class RAGBenchmarkNode(Node):
+    """
+    RAG Performance Benchmarking Node
+
+    Benchmarks RAG systems for performance characteristics.
+
+    When to use:
+    - Best for: System comparison, optimization, capacity planning
+    - Not ideal for: Quality evaluation (use RAGEvaluationNode)
+    - Metrics: Latency, throughput, resource usage, scalability
+
+    Example:
+        benchmark = RAGBenchmarkNode(
+            workload_sizes=[10, 100, 1000],
+            concurrent_users=[1, 5, 10]
+        )
+
+        results = await benchmark.run(
+            rag_systems={"system_a": rag_a, "system_b": rag_b},
+            test_queries=queries
+        )
+
+    Parameters:
+        workload_sizes: Different dataset sizes to test
+        concurrent_users: Concurrency levels to test
+        metrics_interval: How often to collect metrics
+
+    Returns:
+        latency_profiles: Response time distributions
+        throughput_curves: Requests/second at different loads
+        resource_usage: Memory and compute utilization
+        scalability_analysis: How performance scales
+    """
+
+    def __init__(
+        self,
+        name: str = "rag_benchmark",
+        workload_sizes: List[int] = None,
+        concurrent_users: List[int] = None,
+    ):
+        self.workload_sizes = workload_sizes or [10, 100, 1000]
+        self.concurrent_users = concurrent_users or [1, 5, 10]
+        super().__init__(name)
+
+    def get_parameters(self) -> Dict[str, NodeParameter]:
+        return {
+            "rag_systems": NodeParameter(
+                name="rag_systems",
+                type=dict,
+                required=True,
+                description="RAG systems to benchmark",
+            ),
+            "test_queries": NodeParameter(
+                name="test_queries",
+                type=list,
+                required=True,
+                description="Queries for benchmarking",
+            ),
+            "duration": NodeParameter(
+                name="duration",
+                type=int,
+                required=False,
+                default=60,
+                description="Test duration in seconds",
+            ),
+        }
+
+    def run(self, **kwargs) -> Dict[str, Any]:
+        """Run performance benchmarks"""
+        rag_systems = kwargs.get("rag_systems", {})
+        test_queries = kwargs.get("test_queries", [])
+        duration = kwargs.get("duration", 60)
+
+        benchmark_results = {}
+
+        for system_name, system in rag_systems.items():
+            system_results = {
+                "latency_profiles": {},
+                "throughput_curves": {},
+                "resource_usage": {},
+                "scalability_analysis": {},
+            }
+
+            # Test different workload sizes
+            for size in self.workload_sizes:
+                workload = test_queries[:size]
+
+                # Measure latency
+                latencies = []
+                start_time = time.time()
+
+                for query in workload:
+                    query_start = time.time()
+                    # Would call system.run(query=query) in production
+                    # Simulate processing
+                    time.sleep(0.1 + random.random() * 0.1)
+                    latencies.append(time.time() - query_start)
+
+                system_results["latency_profiles"][f"size_{size}"] = {
+                    "p50": statistics.median(latencies),
+                    "p95": sorted(latencies)[int(len(latencies) * 0.95)],
+                    "p99": sorted(latencies)[int(len(latencies) * 0.99)],
+                    "mean": statistics.mean(latencies),
+                    "std_dev": statistics.stdev(latencies) if len(latencies) > 1 else 0,
+                }
+
+                # Calculate throughput
+                total_time = time.time() - start_time
+                throughput = len(workload) / total_time
+                system_results["throughput_curves"][f"size_{size}"] = throughput
+
+            # Test concurrency
+            for users in self.concurrent_users:
+                # Simulate concurrent load
+                concurrent_latencies = []
+
+                # Simplified - would use asyncio/threading in production
+                for _ in range(users * 10):
+                    query_start = time.time()
+                    time.sleep(0.1 + random.random() * 0.2 * users)
+                    concurrent_latencies.append(time.time() - query_start)
+
+                system_results["scalability_analysis"][f"users_{users}"] = {
+                    "avg_latency": statistics.mean(concurrent_latencies),
+                    "throughput_degradation": 1.0 / users,  # Simplified
+                }
+
+            # Simulate resource usage
+            system_results["resource_usage"] = {
+                "memory_mb": 100 + random.randint(0, 500),
+                "cpu_percent": 20 + random.randint(0, 60),
+                "gpu_memory_mb": (
+                    0
+                    if "gpu" not in system_name.lower()
+                    else 1000 + random.randint(0, 3000)
+                ),
+            }
+
+            benchmark_results[system_name] = system_results
+
+        # Comparative analysis
+        comparison = self._compare_systems(benchmark_results)
+
+        return {
+            "benchmark_results": benchmark_results,
+            "comparison": comparison,
+            "test_configuration": {
+                "workload_sizes": self.workload_sizes,
+                "concurrent_users": self.concurrent_users,
+                "duration": duration,
+                "num_queries": len(test_queries),
+            },
+        }
+
+    def _compare_systems(self, results: Dict) -> Dict[str, Any]:
+        """Compare benchmark results across systems"""
+        comparison = {
+            "fastest_system": None,
+            "most_scalable": None,
+            "most_efficient": None,
+            "recommendations": [],
+        }
+
+        # Find fastest system
+        avg_latencies = {}
+        for system, data in results.items():
+            latencies = [v["mean"] for v in data["latency_profiles"].values()]
+            avg_latencies[system] = (
+                statistics.mean(latencies) if latencies else float("inf")
+            )
+
+        comparison["fastest_system"] = min(avg_latencies, key=avg_latencies.get)
+
+        # Find most scalable
+        scalability_scores = {}
+        for system, data in results.items():
+            # Lower degradation = better scalability
+            degradations = [
+                v["throughput_degradation"]
+                for v in data["scalability_analysis"].values()
+            ]
+            scalability_scores[system] = (
+                statistics.mean(degradations) if degradations else 0
+            )
+
+        comparison["most_scalable"] = max(
+            scalability_scores, key=scalability_scores.get
+        )
+
+        # Find most efficient (performance per resource)
+        efficiency_scores = {}
+        for system, data in results.items():
+            throughput = (
+                statistics.mean(data["throughput_curves"].values())
+                if data["throughput_curves"]
+                else 1
+            )
+            memory = data["resource_usage"]["memory_mb"]
+            efficiency_scores[system] = throughput / memory * 1000
+
+        comparison["most_efficient"] = max(efficiency_scores, key=efficiency_scores.get)
+
+        # Generate recommendations
+        comparison["recommendations"] = [
+            f"Use {comparison['fastest_system']} for latency-critical applications",
+            f"Use {comparison['most_scalable']} for high-concurrency scenarios",
+            f"Use {comparison['most_efficient']} for resource-constrained environments",
+        ]
+
+        return comparison
+
+
+@register_node()
+class TestDatasetGeneratorNode(Node):
+    """
+    RAG Test Dataset Generator
+
+    Generates synthetic test datasets for RAG evaluation.
+
+    When to use:
+    - Best for: Creating evaluation benchmarks, testing edge cases
+    - Not ideal for: Production data generation
+    - Output: Queries with ground truth answers and contexts
+
+    Example:
+        generator = TestDatasetGeneratorNode(
+            categories=["factual", "analytical", "comparative"],
+            difficulty_levels=["easy", "medium", "hard"]
+        )
+
+        dataset = generator.run(
+            num_samples=100,
+            domain="machine learning"
+        )
+
+    Parameters:
+        categories: Types of questions to generate
+        difficulty_levels: Complexity levels
+        include_adversarial: Generate tricky cases
+
+    Returns:
+        test_queries: Generated queries with metadata
+        reference_answers: Ground truth answers
+        test_contexts: Relevant documents
+    """
+
+    def __init__(
+        self,
+        name: str = "test_dataset_generator",
+        categories: List[str] = None,
+        include_adversarial: bool = True,
+    ):
+        self.categories = categories or ["factual", "analytical", "comparative"]
+        self.include_adversarial = include_adversarial
+        super().__init__(name)
+
+    def get_parameters(self) -> Dict[str, NodeParameter]:
+        return {
+            "num_samples": NodeParameter(
+                name="num_samples",
+                type=int,
+                required=True,
+                description="Number of test samples",
+            ),
+            "domain": NodeParameter(
+                name="domain",
+                type=str,
+                required=False,
+                default="general",
+                description="Domain for questions",
+            ),
+            "seed": NodeParameter(
+                name="seed",
+                type=int,
+                required=False,
+                description="Random seed for reproducibility",
+            ),
+        }
+
+    def run(self, **kwargs) -> Dict[str, Any]:
+        """Generate test dataset"""
+        num_samples = kwargs.get("num_samples", 10)
+        domain = kwargs.get("domain", "general")
+        seed = kwargs.get("seed")
+
+        if seed:
+            random.seed(seed)
+
+        test_dataset = []
+
+        # Templates for different categories
+        templates = {
+            "factual": [
+                ("What is {concept}?", "Definition and explanation of {concept}"),
+                (
+                    "When was {event} discovered?",
+                    "Discovery date and context of {event}",
+                ),
+                ("Who invented {invention}?", "Inventor and history of {invention}"),
+            ],
+            "analytical": [
+                (
+                    "How does {system} work?",
+                    "Detailed explanation of {system} mechanics",
+                ),
+                (
+                    "What are the advantages of {method}?",
+                    "Benefits and strengths of {method}",
+                ),
+                (
+                    "Why is {principle} important?",
+                    "Significance and applications of {principle}",
+                ),
+            ],
+            "comparative": [
+                (
+                    "Compare {option1} and {option2}",
+                    "Comparison of {option1} vs {option2}",
+                ),
+                (
+                    "What's the difference between {concept1} and {concept2}?",
+                    "Distinctions between concepts",
+                ),
+                (
+                    "Which is better: {choice1} or {choice2}?",
+                    "Trade-offs and recommendations",
+                ),
+            ],
+        }
+
+        # Domain-specific concepts
+        domain_concepts = {
+            "machine learning": [
+                "neural networks",
+                "transformers",
+                "BERT",
+                "attention mechanism",
+                "backpropagation",
+            ],
+            "general": [
+                "democracy",
+                "photosynthesis",
+                "gravity",
+                "internet",
+                "climate change",
+            ],
+        }
+
+        concepts = domain_concepts.get(domain, domain_concepts["general"])
+
+        for i in range(num_samples):
+            category = random.choice(self.categories)
+            template_q, template_a = random.choice(templates[category])
+
+            # Generate specific question
+            if "{concept}" in template_q:
+                concept = random.choice(concepts)
+                query = template_q.format(concept=concept)
+                answer = template_a.format(concept=concept)
+            else:
+                # Handle other placeholders
+                query = template_q
+                answer = template_a
+
+            # Generate contexts
+            contexts = []
+            for j in range(3):
+                contexts.append(
+                    {
+                        "id": f"ctx_{i}_{j}",
+                        "content": f"Context {j+1} about {query}: {answer}",
+                        "relevance": 0.9 - j * 0.1,
+                    }
+                )
+
+            # Add adversarial examples if enabled
+            metadata = {"category": category, "difficulty": "medium"}
+
+            if self.include_adversarial and random.random() < 0.2:
+                # Make it adversarial
+                if random.random() < 0.5:
+                    # Negation
+                    query = f"Is it true that {query.lower()}"
+                    metadata["adversarial_type"] = "negation"
+                else:
+                    # Misleading context
+                    contexts.append(
+                        {
+                            "id": f"ctx_{i}_misleading",
+                            "content": f"Incorrect information: {query} is actually false because...",
+                            "relevance": 0.7,
+                        }
+                    )
+                    metadata["adversarial_type"] = "misleading_context"
+
+            test_dataset.append(
+                {
+                    "id": f"test_{i}",
+                    "query": query,
+                    "reference_answer": answer,
+                    "contexts": contexts,
+                    "metadata": metadata,
+                }
+            )
+
+        return {
+            "test_dataset": test_dataset,
+            "statistics": {
+                "total_samples": len(test_dataset),
+                "category_distribution": {
+                    cat: sum(
+                        1 for t in test_dataset if t["metadata"]["category"] == cat
+                    )
+                    for cat in self.categories
+                },
+                "adversarial_count": sum(
+                    1 for t in test_dataset if "adversarial_type" in t["metadata"]
+                ),
+            },
+            "generation_config": {
+                "domain": domain,
+                "categories": self.categories,
+                "seed": seed,
+            },
+        }
+
+
+# Export all evaluation nodes
+__all__ = ["RAGEvaluationNode", "RAGBenchmarkNode", "TestDatasetGeneratorNode"]