Skip to content

Quiz: Choosing Models and Writing Clear Prompts

Test your understanding of model access, selection, retrieval, and the building blocks of a reliable prompt.

1. What is API access in the context of using a language model?

  1. A chat window in a web browser
  2. A built-in model feature for reasoning through problems
  3. Running a model only on your laptop
  4. The ability to send prompts to a model from a program over the network and receive completions as data
Show Answer

The correct answer is D. API access lets you call a model from a script, pass parameters such as temperature and max tokens, and receive structured responses. A chat window is a graphical front end that itself uses an API, while local models run on your own hardware. Reasoning is a model capability, not an access method.

Concept Tested: API Access

See: Reaching a Model Through an API

2. What is an open-weight model?

  1. A model that is always smaller than hosted models
  2. A model that is free to use through any chat interface
  3. A model whose training data and code are fully public
  4. A model whose trained parameters are published so it can be downloaded and run yourself
Show Answer

The correct answer is D. An open-weight model publishes its trained parameters, making local execution possible. It is not the same as fully open source, because the training data and code may not be released, and it is not necessarily free or small. Open weights are what enable local models.

Concept Tested: Open Weight Model

See: Selecting a Model

3. Which access path is the best fit for classifying 5,000 abstracts overnight and saving the labels to a file?

  1. Chat interface
  2. Local open-weight model on a laptop with no GPU
  3. Hosted API
  4. Emailing the abstracts to a colleague
Show Answer

The correct answer is C. A hosted API is scriptable and designed for batch work over many inputs. A chat interface is impractical at that scale, a weak local machine may be too slow or lack memory for thousands of classifications, and email is not a model access path.

Concept Tested: Model Selection

See: Selecting a Model

4. Why is a local open-weight model especially attractive when results must be reproducible years later?

  1. It always produces the same completion because it is smaller.
  2. It avoids the model version drift that affects hosted models.
  3. It requires no hardware and no setup.
  4. It is always the most capable model available.
Show Answer

The correct answer is B. Hosted models can be updated or retired, so the same prompt may produce different results later. A local open-weight model's weights can be archived with the project and rerun unchanged. It does require hardware, and the strongest hosted models are usually more capable.

Concept Tested: Model Version Drift

See: Selecting a Model

5. What does retrieval-augmented generation (RAG) do?

  1. It retrieves the most relevant passages from your documents and inserts them into the prompt.
  2. It retrains the model on your private documents.
  3. It generates embeddings without using vector search.
  4. It removes the need for any system prompt.
Show Answer

The correct answer is A. RAG embeds your documents, finds the passages most similar to the query by vector search, and places those passages into the prompt so the model answers from grounded text. It does not retrain the model and does not eliminate the system prompt.

Concept Tested: Retrieval Augmented Generation

See: Adding Your Own Documents

  1. It cannot compare vectors mathematically.
  2. It always returns the same number of results, even when the matches are weak.
  3. It only works for text written in English.
  4. It requires retraining the embedding model for every query.
Show Answer

The correct answer is B. Vector search returns the top-k closest vectors regardless of absolute quality, so a weak match still comes back with the same formatting as a strong one. That is why a similarity threshold and human spot-checking are important. Cosine similarity is the comparison method, and the embedding model is not retrained per query.

Concept Tested: Vector Search

See: Adding Your Own Documents

7. In the prompt "Use only information that appears in the attached text," which prompt component is being used?

  1. Role prompting
  2. Output format
  3. Constraint
  4. Few-shot example
Show Answer

The correct answer is C. This sentence limits the sources the model may use, which is a source constraint. It does not assign a role, specify the shape of the output, or provide an example of the desired response.

Concept Tested: Constraints

See: Prompt Engineering

8. What is an output format in prompt engineering?

  1. The set of constraints on what the model may not do
  2. The role assigned to the model
  3. An explicit statement of the shape of the response, such as a table or JSON object
  4. The conversation history included in the prompt
Show Answer

The correct answer is C. An output format tells the model exactly how to structure its answer, making the result parseable and checkable. Constraints limit behavior, role sets the viewpoint, and conversation history is part of the prompt context, not the format.

Concept Tested: Output Format

See: Output Format

9. Which prompt technique is usually most effective when you want the model to learn a format or judgment that is hard to describe in words?

  1. Adding a stronger system prompt
  2. Providing a few worked input-output examples
  3. Raising the temperature
  4. Asking the model to explain its reasoning first
Show Answer

The correct answer is B. Few-shot examples show the model what you want through concrete cases, which is especially useful for subtle formats or classification boundaries. They cost context tokens and may be imitated too closely, but they are the strongest single way to communicate such tasks. Temperature and reasoning prompts do not directly teach format.

Concept Tested: Few-Shot Examples

See: Few-Shot Examples

10. You are asking a model to classify abstracts and report numeric ENOB values. Which addition best guards against invented values?

  1. "If the value is not stated in the abstract, write 'not stated'."
  2. "Act as an expert in analog circuit design."
  3. "Return the answer as a JSON object."
  4. "Provide three examples of classified abstracts."
Show Answer

The correct answer is A. Giving the model a sanctioned way to say "not stated" is an escape constraint and the cheapest defense against hallucinating a number. A role, JSON format, and examples help with other aspects but do not directly prevent invention when the source lacks the value.

Concept Tested: Constraints

See: Constraints