LLM Apps in Production (RAG + Vector DB + Caching) · Lesson

Developing Evaluation Benchmarks

Create custom datasets and benchmarks to systematically test and compare different RAG configurations and improvements.

Developing Evaluation Benchmarks is a free LLM Apps in Production (RAG + Vector DB + Caching) lesson on CoddyKit — lesson 2 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 LLM Apps in Production (RAG + Vector DB + Caching) learning path, one of 4 lessons in the course, and your progress syncs across the web and the CoddyKit app.

Why RAG Benchmarks Matter

Welcome! In this lesson, we'll learn how to create custom evaluation benchmarks for your RAG systems. Benchmarks are like custom test sets that help you measure how well your RAG application performs.

They are crucial for understanding improvements, regressions, and ensuring your RAG system delivers accurate and relevant information to your users.

Custom Benchmarks: The Why

While public datasets like SQuAD or HotpotQA are great for general LLM evaluation, they often don't reflect your specific use case or domain.

Domain Specificity: Your RAG needs to answer questions about your data.
Nuance & Complexity: Public datasets might not capture the unique challenges your users face.
Continuous Improvement: Custom benchmarks allow you to track performance against your evolving needs.

What Makes a RAG Benchmark?

A robust RAG evaluation benchmark typically consists of a few key parts:

Query Set: A collection of representative questions or prompts.
Ground Truth: The "correct" answers or relevant documents for each query.
Evaluation Metrics: The criteria you'll use to measure performance (e.g., accuracy, relevance).

We'll focus on the first two components in this lesson.

Building a Great Query Set

Your query set should mirror the types of questions real users will ask. Think about:

Real User Data: Analyze actual user queries or common support tickets.
Diverse Topics: Cover a wide range of subjects relevant to your RAG's knowledge base.
Varying Difficulty: Include simple, complex, and even ambiguous questions.
Edge Cases: Don't forget queries that might challenge your system.

Example: Query Generation

You can start by manually crafting queries or by using an LLM to generate them based on your documents. Here's a simple Python example of a query set structure:

queries = [
  "What are the benefits of cloud computing?",
  "Explain the capital gains tax in detail.",
  "How do I reset my account password?",
  "What is the company's policy on remote work?",
  "List common cybersecurity threats."
]

for q in queries:
  print(f"Query: {q}")

Establishing Ground Truth

Ground truth is the gold standard against which your RAG's output is measured. For RAG, this often means identifying:

Relevant Documents: Which specific documents should be retrieved for a given query?
Correct Answers: What is the ideal answer based on those documents?

This step often requires human expertise to ensure accuracy.

Structuring Ground Truth

Ground truth can be stored in a structured way, linking queries to their expected relevant context and answers. This allows for automated evaluation.

ground_truth = {
  "What are the benefits of cloud computing?": {
    "relevant_docs": ["doc_cloud_intro.txt", "doc_cloud_benefits.pdf"],
    "answer": "Scalability, cost savings, flexibility, and reliability."
  },
  "How do I reset my account password?": {
    "relevant_docs": ["doc_password_reset_guide.html"],
    "answer": "Go to settings, click 'Forgot Password', and follow the prompts."
  }
}

for query, gt in ground_truth.items():
  print(f"Query: {query}")
  print(f"  Expected Docs: {gt['relevant_docs']}")
  print(f"  Expected Answer: {gt['answer']}\n")

The Human Touch: Annotation

Creating high-quality ground truth often involves human annotation. This means:

Experts Review: Subject matter experts identify relevant documents and craft ideal answers.
Crowdsourcing: For larger datasets, platforms can be used, but quality control is vital.
Consistency: Clear guidelines are essential to ensure annotators label data uniformly.

This ensures your benchmark accurately reflects "correctness."

Benchmarks Evolve

Your RAG system and its data will change over time, and so should your benchmarks! Treat your evaluation benchmarks as living assets:

Add New Queries: Incorporate new user questions or emerging topics.
Update Ground Truth: As your knowledge base grows, update expected answers.
Retire Old Data: Remove outdated information that is no longer relevant.

Regular review keeps your benchmark effective.

Benchmark Essentials

Which of the following are essential components of a robust RAG evaluation benchmark?

Recap: Building Benchmarks

Great job! You've learned how to develop custom evaluation benchmarks for your RAG system. We covered:

The importance of custom, domain-specific benchmarks.
The core components: query sets and ground truth.
Strategies for crafting representative queries and defining accurate ground truth.
The role of human annotation and iterative refinement.

Next, we'll explore how to use these benchmarks to apply key metrics for RAG performance evaluation!

Frequently asked questions

Is the “Developing Evaluation Benchmarks” lesson free?

Yes — the full text of “Developing Evaluation Benchmarks” is free to read here on the web, and the LLM Apps in Production (RAG + Vector DB + Caching) 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 LLM Apps in Production (RAG + Vector DB + Caching) course, upgrade to CoddyKit PRO.

What will I learn in “Developing Evaluation Benchmarks”?

Create custom datasets and benchmarks to systematically test and compare different RAG configurations and improvements. You practise LLM Apps in Production (RAG + Vector DB + Caching) 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 LLM Apps in Production (RAG + Vector DB + Caching)?

No prior experience is required. LLM Apps in Production (RAG + Vector DB + Caching) on CoddyKit is structured for beginners through advanced learners; this is — lesson 2 of 4, so you can start here or from the beginning and move at your own pace.

How long does the “Developing Evaluation Benchmarks” 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 LLM Apps in Production (RAG + Vector DB + Caching) lesson?

Yes. Every LLM Apps in Production (RAG + Vector DB + Caching) 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 LLM Apps in Production (RAG + Vector DB + Caching)