Why RAG gives agents knowledge in Agentic AI - Why Metrics Matter

Metrics & Evaluation - Why RAG gives agents knowledge

Which metric matters for this concept and WHY

For Retrieval-Augmented Generation (RAG) agents, the key metric is retrieval accuracy. This measures how well the agent finds the right information from its knowledge base. Good retrieval accuracy means the agent uses correct facts to answer questions. Another important metric is response relevance, which checks if the agent's answers are useful and on-topic. These metrics matter because RAG agents combine retrieved knowledge with language generation to give informed answers.

Confusion matrix or equivalent visualization (ASCII)

    Retrieved Info vs. Correct Info

           | Correct Info Present | Correct Info Absent
    -------|----------------------|-------------------
    Retrieved     |        TP           |        FP         
    Not Retrieved |        FN           |        TN         

    TP = Correct info retrieved
    FP = Wrong info retrieved
    FN = Correct info missed
    TN = Correctly ignored wrong info
    

This matrix helps measure retrieval precision and recall, which affect the agent's knowledge quality.

Precision vs Recall tradeoff with concrete examples

Precision means how many retrieved facts are actually correct. High precision means the agent rarely uses wrong info. For example, a medical assistant agent must have high precision to avoid giving harmful advice.

Recall means how many correct facts the agent finds out of all possible correct facts. High recall means the agent finds most relevant info. For example, a research assistant agent benefits from high recall to gather all useful data.

RAG agents balance precision and recall to give knowledgeable and trustworthy answers.

What "good" vs "bad" metric values look like for this use case

• Good: Retrieval precision and recall above 85%, leading to accurate and complete answers.
• Bad: Precision below 50% means many wrong facts used, causing misinformation.
• Bad: Recall below 40% means many relevant facts missed, leading to incomplete answers.
• Balanced high precision and recall ensure the agent's knowledge is reliable and helpful.

Metrics pitfalls

• Accuracy paradox: High overall accuracy can hide poor retrieval if most queries are easy.
• Data leakage: If the agent's knowledge base contains test answers, metrics will be unrealistically high.
• Overfitting: The agent may memorize facts but fail to generalize to new questions.
• Ignoring relevance: Retrieving many facts but not relevant ones inflates recall but hurts usefulness.

Self-check

Your RAG agent has 98% retrieval accuracy but only 12% recall on key facts. Is it good for production? Why not?

Answer: No, because the agent misses most important facts (low recall). It may give precise but incomplete answers, which can mislead users. Improving recall is critical for trustworthy knowledge.

Key Result

Retrieval precision and recall are key to RAG agents' knowledge quality; both must be balanced for reliable answers.

Practice

1. What is the main reason RAG (Retrieval-Augmented Generation) helps AI agents have better knowledge?

easy

A. It ignores external information sources.

B. It only uses pre-trained data without updates.

C. It combines retrieving information with generating answers.

D. It relies solely on random guessing.

Solution

1. Step 1: Understand RAG's components

   RAG combines two parts: retrieval (finding relevant info) and generation (creating answers).

2. Step 2: Connect combination to knowledge improvement

   By mixing retrieval and generation, agents can use both stored and new info, improving knowledge.

3. Final Answer:

   It combines retrieving information with generating answers. -> Option C

4. Quick Check:

   RAG = retrieval + generation [OK]

Hint: Remember RAG mixes retrieval and generation [OK]

Common Mistakes:

• Thinking RAG only uses pre-trained data
• Believing RAG ignores external info
• Assuming RAG guesses randomly

2. Which of the following is the correct way to describe RAG's process in simple terms?

easy

A. RAG retrieves relevant documents, then generates answers using them.

B. RAG generates answers first, then searches for info.

C. RAG only retrieves documents without generating answers.

D. RAG randomly selects answers without retrieval.

Solution

1. Step 1: Identify RAG's sequence

   RAG first retrieves relevant documents from a source.

2. Step 2: Understand generation step

   Then it generates answers based on the retrieved documents.

3. Final Answer:

   RAG retrieves relevant documents, then generates answers using them. -> Option A

4. Quick Check:

   Retrieve then generate [OK]

Hint: RAG retrieves first, then generates answers [OK]

Common Mistakes:

• Thinking generation happens before retrieval
• Believing RAG only retrieves without generation
• Assuming random answer selection

3. Given this simplified code snippet for a RAG agent:

retrieved_docs = ['Doc about cats', 'Doc about dogs']
query = 'Tell me about cats'
answer = generate_answer(query, retrieved_docs)
print(answer)

What is the expected output behavior?

medium

A. The answer will only use the query without documents.

B. The answer will ignore retrieved_docs and be random.

C. The code will cause an error because generate_answer is undefined.

D. The answer will be generated using information about cats and dogs.

Solution

1. Step 1: Understand inputs to generate_answer

   The function gets the query and the retrieved documents about cats and dogs.

2. Step 2: Predict output behavior

   Since retrieved_docs include relevant info, the answer will use that info to respond about cats.

3. Final Answer:

   The answer will be generated using information about cats and dogs. -> Option D

4. Quick Check:

   RAG uses retrieved docs to generate answers [OK]

Hint: Check if retrieved docs are used in generation [OK]

Common Mistakes:

• Assuming generate_answer is undefined error
• Thinking answer ignores retrieved docs
• Believing answer is random

4. Consider this code snippet for a RAG agent:

def rag_agent(query):
    docs = retrieve_docs(query)
    answer = generate_answer(docs)
    return answer

print(rag_agent('What is AI?'))

What is the main error in this code?

medium

A. generate_answer is called without the query parameter.

B. retrieve_docs is missing the query argument.

C. rag_agent returns docs instead of answer.

D. print statement is outside the function.

Solution

1. Step 1: Check function calls and parameters

   retrieve_docs is called with query, which is correct.

2. Step 2: Identify generate_answer call issue

   generate_answer is called with only docs, but it needs both query and docs to generate a proper answer.

3. Final Answer:

   generate_answer is called without the query parameter. -> Option A

4. Quick Check:

   generate_answer needs query and docs [OK]

Hint: Check if all required parameters are passed to functions [OK]

Common Mistakes:

• Thinking retrieve_docs lacks argument
• Believing rag_agent returns wrong value
• Confusing print statement placement

5. How does RAG improve an AI agent's ability to answer questions about recent events not in its training data?

hard

A. By only relying on its fixed training data without updates.

B. By retrieving up-to-date documents and generating answers using them.

C. By guessing answers based on old data patterns.

D. By ignoring external information and focusing on generation.

Solution

1. Step 1: Understand RAG's retrieval role

   RAG retrieves current documents from external sources, including recent events.

2. Step 2: Understand generation with new info

   It then generates answers using this fresh info, allowing it to handle new questions accurately.

3. Final Answer:

   By retrieving up-to-date documents and generating answers using them. -> Option B

4. Quick Check:

   RAG uses fresh retrieval for new knowledge [OK]

Hint: Remember RAG updates knowledge via retrieval [OK]

Common Mistakes:

• Thinking RAG only uses old training data
• Assuming RAG guesses without info
• Believing RAG ignores external data