What is faithfulness metric in RAG evaluation
faithfulness metric in Retrieval-Augmented Generation (RAG) evaluation measures how accurately a generated answer reflects the content of the retrieved documents without hallucination. It ensures the language model's output is grounded and verifiable against the source knowledge base.Faithfulness metric is an evaluation measure that quantifies how well a RAG model's generated response is supported by the retrieved evidence, ensuring the answer is truthful and grounded.How it works
The faithfulness metric assesses whether the output of a Retrieval-Augmented Generation (RAG) system accurately reflects the information contained in the retrieved documents. Imagine a student answering a question by quoting a textbook: faithfulness checks if the student's answer truly matches the textbook content rather than inventing facts. This is crucial because RAG models combine retrieval with generation, so the generated text should be grounded in the retrieved knowledge, not hallucinated.
Mechanically, faithfulness is often measured by comparing the generated answer to the retrieved passages using overlap metrics (like ROUGE or BLEU), entailment checks with natural language inference models, or human annotation verifying factual consistency.
Concrete example
Here is a simplified Python example using a natural language inference (NLI) model to estimate faithfulness by checking if the generated answer is entailed by the retrieved document text.
from transformers import pipeline
import os
# Load an NLI model for entailment checking
nli = pipeline('text-classification', model='facebook/bart-large-mnli')
retrieved_doc = "The Eiffel Tower is located in Paris and was completed in 1889."
generated_answer = "The Eiffel Tower is in Paris."
# Prepare premise and hypothesis for entailment
premise = retrieved_doc
hypothesis = generated_answer
result = nli(f'{premise} </s></s> {hypothesis}')
# Check if the answer is entailed by the document
is_faithful = any(label == 'ENTAILMENT' and score > 0.9 for label, score in [(r['label'], r['score']) for r in result])
print(f'Faithfulness: {is_faithful}')Faithfulness: True
When to use it
Use the faithfulness metric when evaluating RAG systems that generate answers based on retrieved documents, especially in domains requiring factual accuracy like healthcare, law, or scientific research. It helps ensure the model does not hallucinate or fabricate unsupported information.
Do not rely solely on faithfulness metrics when the retrieval quality is poor or when the task requires creative generation beyond factual grounding.
Key Takeaways
- Faithfulness measures how well a RAG model's output is supported by retrieved documents.
- It prevents hallucination by ensuring generated answers are grounded in source data.
- Natural language inference models are commonly used to automate faithfulness evaluation.
- Use faithfulness metrics in factual domains where accuracy is critical.
- Faithfulness complements but does not replace retrieval quality evaluation.