RAG 성능의 핵심 지표

When building Retrieval Augmented Generation (RAG) systems, it's not enough to just deploy them. We need to know if they're actually working well!

Evaluating RAG is more complex than evaluating a standalone Large Language Model (LLM) because it involves two main stages: retrieval and generation.

Traditional LLM evaluation metrics often focus on the quality of generated text, like fluency or coherence. But RAG systems have specific goals:

• To provide answers grounded in facts.
• To avoid 'hallucinations' (making up information).
• To use only relevant information from your data.

This requires a special set of metrics.

The first key metric is Context Relevance.

• It measures how pertinent the retrieved documents or 'context' are to the user's original question.
• If the retriever fetches irrelevant information, the LLM won't have good material to work with, leading to poor answers.

High context relevance means your retriever is doing its job well!

Let's say a user asks: "What are the benefits of eating apples?"

Good Context: "Apples are rich in fiber, vitamin C, and antioxidants..." (High relevance)

Poor Context: "Oranges are citrus fruits. Apples can be green or red..." (Low relevance for 'benefits')

The quality of the retrieved context directly impacts the LLM's ability to answer correctly.

Next, we have Faithfulness (also called 'groundedness').

• This metric checks if the LLM's generated answer is entirely supported by the retrieved context.
• It's crucial for preventing 'hallucinations' – where the LLM invents facts not present in the source material.

A faithful RAG system will only provide information it can verify from its sources.

User Question: "What is the capital of France?"

Retrieved Context: "Paris is the capital of France, known for the Eiffel Tower."

Faithful Answer: "The capital of France is Paris." (Supported by context)

Unfaithful Answer: "The capital of France is Lyon, a beautiful city." (Not supported by context)

Answer Relevance evaluates whether the LLM's generated response directly addresses the user's original question.

• Even if the answer is faithful and based on relevant context, it might still be too verbose, tangential, or miss the point of the question.
• This metric ensures the final output is useful and to-the-point for the user.

User Question: "When was the internet invented?"

Retrieved Context: "The internet's origins trace back to the 1960s with ARPANET..."

Relevant Answer: "The internet's origins trace back to the 1960s with ARPANET."

Irrelevant Answer: "The internet is a global network of computers. It has revolutionized communication." (Doesn't answer 'when')

While the previous metrics evaluate the RAG system holistically, classic Information Retrieval (IR) metrics like Precision and Recall are vital for the retrieval component.

• Precision: What percentage of the retrieved documents are actually relevant? (Minimize irrelevant documents)
• Recall: What percentage of all truly relevant documents were actually retrieved? (Minimize missed relevant documents)

Balancing these two is key for feeding the LLM the best possible context.

A RAG system retrieves documents, then generates an answer. Consider the following scenario:

User Question: "What is the typical lifespan of a domestic cat?"

Retrieved Context: "Domestic cats usually live for 12 to 18 years. Some can live longer."

LLM Answer: "Cats are furry animals that enjoy sleeping and playing. Their lifespan varies."

Congratulations! You've learned about the critical metrics for evaluating RAG systems:

• Context Relevance: How good is the retrieved information?
• Faithfulness: Is the answer true to the retrieved context?
• Answer Relevance: Does the answer address the user's question?
• Precision & Recall: How effective is the retrieval component?

Mastering these helps you build more accurate, reliable, and useful RAG applications.