LangChain / RAG / Vector DBs · Lesson

Evaluating RAG System Performance

Learn metrics and techniques to assess the accuracy, relevance, and coherence of your RAG application's responses.

Evaluating RAG System Performance is a free LangChain / RAG / Vector DBs lesson on CoddyKit — lesson 3 of 4. You can read the complete lesson below for free — then practise it hands-on in the browser with a built-in code editor and a 24/7 AI tutor. It is part of the LangChain / RAG / Vector DBs learning path, one of 4 lessons in the course, and your progress syncs across the web and the CoddyKit app.

Why Evaluate RAG Systems?

When you build a Retrieval Augmented Generation (RAG) application, it's crucial to know if it's working as intended. Evaluation helps us understand if our system is providing accurate, relevant, and helpful answers.

Without proper evaluation, it's hard to tell if changes to your RAG pipeline (like new embedding models or chunking strategies) are actually making it better or worse.

Key Aspects of RAG Evaluation

Evaluating a RAG system involves looking at different aspects. We generally focus on:

Retrieval Quality: Are the right documents being found?
Generation Quality: Is the LLM producing good answers based on those documents?

These two parts are often evaluated separately and then together to get a full picture.

The Importance of Ground Truth

To evaluate effectively, especially for automated metrics, you need a 'ground truth' dataset. This means having:

A set of user queries.
The truly relevant documents for each query.
The ideal, correct answer for each query.

This reference data allows us to compare our RAG system's outputs against what's considered correct.

Retrieval Metrics: Precision & Recall

For the retrieval component, we often use metrics like Precision and Recall.

Precision: Out of all the documents retrieved, how many were actually relevant? High precision means fewer irrelevant documents.
Recall: Out of all the truly relevant documents, how many did our system manage to retrieve? High recall means fewer missed relevant documents.

These help us gauge how good our document search is.

Code: Calculating Retrieval Metrics

This Python snippet demonstrates how to calculate precision and recall for a mock retrieval scenario. relevant_docs are the 'ground truth' and retrieved_docs are what our system found.

def evaluate_retrieval(relevant_docs, retrieved_docs):
    true_positives = len(relevant_docs.intersection(retrieved_docs))

    precision = true_positives / len(retrieved_docs) if len(retrieved_docs) > 0 else 0
    recall = true_positives / len(relevant_docs) if len(relevant_docs) > 0 else 0

    print(f"Precision: {precision:.2f}")
    print(f"Recall: {recall:.2f}")

if __name__ == "__main__":
    # Example 1: Perfect retrieval
    relevant_set_1 = {"docA", "docB", "docC"}
    retrieved_set_1 = {"docA", "docB", "docC"}
    print("--- Example 1: Perfect Retrieval ---")
    evaluate_retrieval(relevant_set_1, retrieved_set_1)

    # Example 2: Some missing, some irrelevant
    relevant_set_2 = {"docX", "docY", "docZ"}
    retrieved_set_2 = {"docX", "docA", "docY"}
    print("\n--- Example 2: Mixed Retrieval ---")
    evaluate_retrieval(relevant_set_2, retrieved_set_2)

Generation Metrics: Faithfulness

For the generated answer, faithfulness (also called 'factuality') is key. It measures whether the LLM's answer is truly supported by the retrieved documents.

A RAG system should not 'hallucinate' or make up information. If a fact isn't in the retrieved context, the LLM shouldn't include it in its answer.

Generation Metrics: Answer Relevance

Answer relevance assesses if the generated answer directly addresses the user's original query. An answer might be faithful to the retrieved documents but still not relevant if the documents themselves were off-topic.

This metric helps ensure the RAG system stays focused on the user's actual information need.

Human vs. Automated Evaluation

While automated metrics are efficient, human evaluation is often essential for subjective qualities like fluency, tone, and nuanced relevance.

Automated: Fast, scalable, good for objective metrics (precision, recall, some faithfulness checks).
Human: Gold standard for subjective quality, crucial for complex queries, but expensive and slow.

A hybrid approach, using both, is often best for comprehensive RAG evaluation.

Tools for RAG Evaluation

Several libraries and frameworks help streamline RAG evaluation:

LangChain: Provides built-in evaluation modules, including reference-free metrics and integrations with LLM-as-a-judge.
Ragas: Specifically designed for RAG evaluation, offering metrics like faithfulness, answer relevance, context precision, and context recall.
TruLens: Offers observability and evaluation for LLM applications, including RAG, with deep insights into chain execution.

These tools automate much of the metric calculation and reporting.

Quick Check: RAG Metrics

Which of the following are key metrics used to evaluate the quality of the generated answer in a RAG system?

Recap: Evaluating RAG Performance

We've learned that evaluating RAG systems is vital for ensuring they deliver accurate and relevant information. Key takeaways include:

Evaluation covers both retrieval (finding relevant docs) and generation (creating good answers).
Ground truth data is essential for reliable evaluation.
Metrics like Precision and Recall measure retrieval quality.
Faithfulness and Answer Relevance assess the quality of the generated response.
A combination of automated and human evaluation provides the most comprehensive insights.
Tools like LangChain, Ragas, and TruLens can assist in setting up robust evaluation pipelines.

Frequently asked questions

Is the “Evaluating RAG System Performance” lesson free?

Yes — the full text of “Evaluating RAG System Performance” is free to read here on the web, and the LangChain / RAG / Vector DBs course includes 4 lessons in total. To practise it interactively (a built-in code editor and a 24/7 AI tutor) and unlock the rest of the LangChain / RAG / Vector DBs course, upgrade to CoddyKit PRO.

What will I learn in “Evaluating RAG System Performance”?

Learn metrics and techniques to assess the accuracy, relevance, and coherence of your RAG application's responses. You practise LangChain / RAG / Vector DBs with hands-on code you run directly in the browser, and a 24/7 AI tutor answers your questions as you work through the lesson.

Do I need any experience to start LangChain / RAG / Vector DBs?

No prior experience is required. LangChain / RAG / Vector DBs on CoddyKit is structured for beginners through advanced learners; this is — lesson 3 of 4, so you can start here or from the beginning and move at your own pace.

How long does the “Evaluating RAG System Performance” lesson take?

Most CoddyKit lessons take about 5–10 minutes. Each one is bite-sized and interactive, so you make steady progress and pick up exactly where you left off across the web and the app.

Can I write and run code in this LangChain / RAG / Vector DBs lesson?

Yes. Every LangChain / RAG / Vector DBs lesson includes a built-in code editor, so you write and run real code right in your browser and get instant AI feedback — no local setup required.

All lessons in this course

← Back to LangChain / RAG / Vector DBs