Combining retrieved context with LLM in Prompt Engineering / GenAI - Model Metrics & Evaluation

Metrics & Evaluation - Combining retrieved context with LLM

Which metric matters for this concept and WHY

When combining retrieved context with a large language model (LLM), the key metric is accuracy or relevance of the model's output. This is because the goal is to produce answers that correctly use the retrieved information. Metrics like precision and recall help measure how well the model uses the right context without adding wrong or irrelevant details.

For example, if the model retrieves documents to answer a question, precision measures how many retrieved facts are actually correct in the answer, while recall measures how many correct facts from the documents are included. Balancing these ensures the LLM output is both accurate and complete.

Confusion matrix or equivalent visualization (ASCII)

Confusion Matrix for context usage in LLM output:

               | Predicted Relevant | Predicted Irrelevant |
---------------|--------------------|----------------------|
Actually Relevant |        TP = 80      |        FN = 20       |
Actually Irrelevant |       FP = 10      |        TN = 90       |

Total samples = 80 + 20 + 10 + 90 = 200

Calculations:
Precision = TP / (TP + FP) = 80 / (80 + 10) = 0.89
Recall = TP / (TP + FN) = 80 / (80 + 20) = 0.80
F1 Score = 2 * (0.89 * 0.80) / (0.89 + 0.80) ≈ 0.84

This shows the model retrieves mostly relevant context (high precision) and covers most relevant facts (good recall).

Precision vs Recall tradeoff with concrete examples

When combining retrieved context with an LLM, there is a tradeoff between precision and recall:

• High precision, low recall: The model uses only very certain retrieved facts. This means fewer wrong details but may miss some important information. Good when you want very trustworthy answers.
• High recall, low precision: The model tries to include all possible relevant facts, even if some are uncertain. This covers more information but risks adding wrong or irrelevant details. Useful when completeness is critical.

Example: For a medical question, high precision is important to avoid wrong advice. For a research summary, high recall helps include all relevant studies.

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

Good metrics:

• Precision ≥ 0.85: Most retrieved context used is correct.
• Recall ≥ 0.75: Most relevant context is included in the output.
• F1 Score ≥ 0.80: Balanced precision and recall.

Bad metrics:

• Precision < 0.5: Many irrelevant or wrong facts included.
• Recall < 0.4: Many relevant facts missed.
• F1 Score < 0.5: Poor balance, unreliable output.

These values depend on the application but generally, higher precision and recall mean better use of retrieved context with the LLM.

Metrics pitfalls

• Accuracy paradox: High overall accuracy can be misleading if the dataset is imbalanced (e.g., many irrelevant facts). Precision and recall give a clearer picture.
• Data leakage: If the LLM sees the answer during training, metrics will be unrealistically high.
• Overfitting: The model may memorize retrieved context instead of understanding it, inflating precision but hurting generalization.
• Ignoring context quality: Metrics assume retrieved context is correct; poor retrieval hurts final output regardless of LLM quality.

Self-check question

Your model combining retrieved context with an LLM has 98% accuracy but only 12% recall on relevant facts. Is it good for production? Why or why not?

Answer: No, it is not good. The very low recall means the model misses most relevant facts, so the output is incomplete. Even though accuracy is high, the model fails to use enough correct context, which can lead to poor or misleading answers.

Key Result

Precision and recall are key to measuring how well the LLM uses retrieved context; balanced high values indicate good model performance.

Practice

1. Why do we combine retrieved context with a large language model (LLM)?

easy

A. To give the model extra information it did not learn before

B. To make the model run faster

C. To reduce the size of the model

D. To replace the model's training data

Solution

1. Step 1: Understand the purpose of retrieved context

   Retrieved context provides additional information that the model might not have seen during training.

2. Step 2: Connect context to model output quality

   Providing this extra information helps the model give better and more accurate answers.

3. Final Answer:

   To give the model extra information it did not learn before -> Option A

4. Quick Check:

   Extra info improves answers = D [OK]

Hint: Extra info helps model answer better [OK]

Common Mistakes:

• Thinking context speeds up the model
• Believing context shrinks the model size
• Assuming context replaces training data

2. Which of the following is the correct way to combine retrieved context with an LLM prompt?

easy

A. prompt = question * context

B. prompt = question + context

C. prompt = context + ' ' + question

D. prompt = context - question

Solution

1. Step 1: Understand prompt construction

   The prompt should start with the context followed by the question to give the model relevant info first.

2. Step 2: Check syntax correctness

   Using string concatenation with '+' is correct; multiplication or subtraction of strings is invalid.

3. Final Answer:

   prompt = context + ' ' + question -> Option C

4. Quick Check:

   Context before question with '+' = A [OK]

Hint: Concatenate context and question with + [OK]

Common Mistakes:

• Putting question before context
• Using * or - operators on strings
• Not adding space between context and question

3. Given the code below, what will be the output?

context = 'The capital of France is Paris.'
question = 'What is the capital of France?'
prompt = context + ' ' + question
response = llm.generate(prompt)
print(response)

Assuming llm.generate() returns the model's answer, what is the likely output?

medium

A. Paris

B. London

C. I don't know

D. Error: undefined variable

Solution

1. Step 1: Analyze the prompt content

   The prompt includes the context 'The capital of France is Paris.' followed by the question.

2. Step 2: Predict model output based on context

   The model uses the context to answer correctly with 'Paris'.

3. Final Answer:

   Paris -> Option A

4. Quick Check:

   Context guides answer = Paris [OK]

Hint: Context gives correct answer to question [OK]

Common Mistakes:

• Ignoring context and guessing wrong
• Assuming code error without cause
• Thinking model says 'I don't know'

4. You wrote this code to combine context with a question:

context = 'Water boils at 100 degrees Celsius.'
question = 'At what temperature does water boil?'
prompt = question + ' ' + context
response = llm.generate(prompt)
print(response)

Why might the model give a less accurate answer?

medium

A. Because the context is missing important info

B. Because the question comes before the context, confusing the model

C. Because the model cannot handle string concatenation

D. Because the prompt is too short

Solution

1. Step 1: Check prompt order

   The prompt puts the question before the context, which may confuse the model about what info to use.

2. Step 2: Understand best practice

   Context should come first to provide relevant info before the question.

3. Final Answer:

   Because the question comes before the context, confusing the model -> Option B

4. Quick Check:

   Context before question improves accuracy = B [OK]

Hint: Put context before question in prompt [OK]

Common Mistakes:

• Thinking model can't concatenate strings
• Assuming context lacks info
• Believing prompt length is the issue

5. You want to build a system that answers questions about a company's products using an LLM. You have a large product manual. What is the best way to combine the manual with the LLM to get accurate answers?

hard

A. Train a new LLM from scratch on the manual

B. Feed the entire manual as a prompt to the LLM every time

C. Only ask the question without any manual context

D. Retrieve relevant sections from the manual and add them as context before the question in the prompt

Solution

1. Step 1: Consider prompt size limits

   Feeding the entire manual is too large and inefficient for the LLM prompt.

2. Step 2: Use retrieval to select relevant info

   Retrieving relevant sections and adding them as context helps the model answer accurately without overload.

3. Step 3: Evaluate other options

   Asking without context misses info; training new LLM is costly and unnecessary.

4. Final Answer:

   Retrieve relevant sections from the manual and add them as context before the question in the prompt -> Option D

5. Quick Check:

   Relevant context retrieval + LLM = A [OK]

Hint: Retrieve relevant info, then prompt LLM [OK]

Common Mistakes:

• Trying to input entire manual at once
• Ignoring context and asking only question
• Thinking retraining is always needed