This code demonstrates the use of the deepeval library to perform comprehensive evaluations of Retrieval-Augmented Generation (RAG) systems. It covers various evaluation metrics and provides a framework for creating and running test cases.
Key Components
- Correctness Evaluation
- Faithfulness Evaluation
- Contextual Relevancy Evaluation
- Combined Evaluation of Multiple Metrics
- Batch Test Case Creation
Evaluation Metrics
1. Correctness (GEval)
- Evaluates whether the actual output is factually correct based on the expected output.
- Uses GPT-4 as the evaluation model.
- Compares the expected and actual outputs.
2. Faithfulness (FaithfulnessMetric)
- Assesses whether the generated answer is faithful to the provided context.
- Uses GPT-4 as the evaluation model.
- Can provide detailed reasons for the evaluation.
3. Contextual Relevancy (ContextualRelevancyMetric)
- Evaluates how relevant the retrieved context is to the question and answer.
- Uses GPT-4 as the evaluation model.
- Can provide detailed reasons for the evaluation.
Key Features
- Flexible Metric Configuration: Each metric can be customized with different models and parameters.
- Multi-Metric Evaluation: Ability to evaluate test cases using multiple metrics simultaneously.
- Batch Test Case Creation: Utility function to create multiple test cases efficiently.
- Detailed Feedback: Options to include detailed reasons for evaluation results.
Benefits of this Approach
- Comprehensive Evaluation: Covers multiple aspects of RAG system performance.
- Flexibility: Easy to add or modify evaluation metrics and test cases.
- Scalability: Capable of handling multiple test cases and metrics efficiently.
- Interpretability: Provides detailed reasons for evaluation results, aiding in system improvement.
This deep evaluation approach using the deepeval library offers a robust framework for assessing the performance of RAG systems. By evaluating correctness, faithfulness, and contextual relevancy, it provides a multi-faceted view of system performance. This comprehensive evaluation is crucial for identifying areas of improvement and ensuring the reliability and effectiveness of RAG systems in real-world applications.
In [18]:
from deepeval import evaluate
from deepeval.metrics import GEval, FaithfulnessMetric, ContextualRelevancyMetric
from deepeval.test_case import LLMTestCase, LLMTestCaseParamsTest Correctness
In [ ]:
correctness_metric = GEval(
name="Correctness",
model="gpt-4o",
evaluation_params=[
LLMTestCaseParams.EXPECTED_OUTPUT,
LLMTestCaseParams.ACTUAL_OUTPUT],
evaluation_steps=[
"Determine whether the actual output is factually correct based on the expected output."
],
)
gt_answer = "Madrid is the capital of Spain."
pred_answer = "MadriD."
test_case_correctness = LLMTestCase(
input="What is the capital of Spain?",
expected_output=gt_answer,
actual_output=pred_answer,
)
correctness_metric.measure(test_case_correctness)
print(correctness_metric.score)Test faithfulness
In [ ]:
question = "what is 3+3?"
context = ["6"]
generated_answer = "6"
faithfulness_metric = FaithfulnessMetric(
threshold=0.7,
model="gpt-4",
include_reason=False
)
test_case = LLMTestCase(
input = question,
actual_output=generated_answer,
retrieval_context=context
)
faithfulness_metric.measure(test_case)
print(faithfulness_metric.score)
print(faithfulness_metric.reason)Test contextual relevancy
In [ ]:
actual_output = "then go somewhere else."
retrieval_context = ["this is a test context","mike is a cat","if the shoes don't fit, then go somewhere else."]
gt_answer = "if the shoes don't fit, then go somewhere else."
relevance_metric = ContextualRelevancyMetric(
threshold=1,
model="gpt-4",
include_reason=True
)
relevance_test_case = LLMTestCase(
input="What if these shoes don't fit?",
actual_output=actual_output,
retrieval_context=retrieval_context,
expected_output=gt_answer,
)
relevance_metric.measure(relevance_test_case)
print(relevance_metric.score)
print(relevance_metric.reason)In [34]:
new_test_case = LLMTestCase(
input="What is the capital of Spain?",
expected_output="Madrid is the capital of Spain.",
actual_output="MadriD.",
retrieval_context=["Madrid is the capital of Spain."]
)Test two different cases together with several metrics together
In [ ]:
evaluate(
test_cases=[relevance_test_case, new_test_case],
metrics=[correctness_metric, faithfulness_metric, relevance_metric]
)Funcion to create multiple LLMTestCases based on four lists:
- Questions
- Ground Truth Answers
- Generated Answers
- Retrieved Documents – Each element is a list
In [ ]:
def create_deep_eval_test_cases(questions, gt_answers, generated_answers, retrieved_documents):
return [
LLMTestCase(
input=question,
expected_output=gt_answer,
actual_output=generated_answer,
retrieval_context=retrieved_document
)
for question, gt_answer, generated_answer, retrieved_document in zip(
questions, gt_answers, generated_answers, retrieved_documents
)
]
