What is faithfulness in RAG evaluation
Retrieval-Augmented Generation (RAG) evaluation, faithfulness measures how accurately the generated response reflects and is supported by the retrieved source documents. It ensures the output is grounded in the evidence provided by the retrieval system rather than hallucinated or fabricated.Faithfulness in Retrieval-Augmented Generation (RAG) evaluation is the degree to which generated answers accurately reflect and are supported by the retrieved source documents.How it works
Faithfulness in RAG evaluation checks if the language model's generated answer truly aligns with the content of the retrieved documents. Imagine a student answering a question by quoting a textbook: faithfulness means the student's answer matches the textbook facts without adding false info. In RAG, the retrieval system fetches relevant documents, and the language model generates an answer based on them. Faithfulness ensures the answer is grounded in those documents, not invented.
Concrete example
Suppose a RAG system retrieves two documents about the Eiffel Tower:
- Doc 1: "The Eiffel Tower is 324 meters tall."
- Doc 2: "It was completed in 1889 for the Paris Exposition."
If the generated answer is "The Eiffel Tower is 324 meters tall and was completed in 1889," it is faithful because it matches the retrieved facts. If it says "The Eiffel Tower is 500 meters tall," it is unfaithful because it contradicts the source.
from openai import OpenAI
import os
client = OpenAI(api_key=os.environ["OPENAI_API_KEY"])
retrieved_docs = [
"The Eiffel Tower is 324 meters tall.",
"It was completed in 1889 for the Paris Exposition."
]
prompt = f"Based on these documents, answer: How tall is the Eiffel Tower and when was it completed?\nDocuments:\n" + "\n".join(retrieved_docs)
response = client.chat.completions.create(
model="gpt-4o",
messages=[{"role": "user", "content": prompt}]
)
print(response.choices[0].message.content)The Eiffel Tower is 324 meters tall and was completed in 1889 for the Paris Exposition.
When to use it
Use faithfulness evaluation when deploying RAG systems in domains requiring high factual accuracy and trust, such as healthcare, legal, or scientific research. It prevents hallucinations by verifying that generated answers are supported by retrieved evidence. Avoid relying solely on faithfulness metrics when the retrieval quality is poor or when creative generation is acceptable.
Key terms
| Term | Definition |
|---|---|
| Faithfulness | Degree to which generated output accurately reflects retrieved source documents. |
| Retrieval-Augmented Generation (RAG) | AI architecture combining document retrieval with language model generation. |
| Hallucination | When a model generates information not supported by any source or fact. |
| Retrieval system | Component that fetches relevant documents or knowledge for the language model. |
Key Takeaways
- Faithfulness ensures RAG outputs are grounded in retrieved documents, reducing hallucinations.
- Evaluate faithfulness by comparing generated answers against source document facts.
- Use faithfulness metrics in critical domains requiring factual accuracy and trust.
- Poor retrieval quality undermines faithfulness, so retrieval and generation must both be strong.