← Back to Week 11

Week 11: Practical LLM Integration & API Development

Knowledge Check — Select an answer to see immediate feedback.

Questions 10

Question 1

When configuring an LLM API call, a temperature value of around 0.2–0.4 is most appropriate for which type of task?

A) Brainstorming creative story ideas where diversity of output is desired
B) Generating focused, deterministic responses such as structured data extraction or code generation
C) Maximizing the length of responses regardless of content
D) Reducing the cost of an API call by limiting the number of tokens generated

Question 2

System prompts are a fundamental tool for controlling LLM behavior. Which of the following best describes their role?

A) They inject examples of desired outputs directly into the model's weights through fine-tuning
B) They set the context and behavioral parameters — including role, constraints, and output format — at the start of the interaction
C) They are used exclusively to limit the model to a specific topic and will cause it to refuse all other queries
D) They replace the need for few-shot examples because they contain implicit demonstrations

Question 3

Chain-of-thought (CoT) prompting improves model performance on complex tasks primarily by doing what?

A) Reducing the number of tokens the model needs to generate by compressing its reasoning into a single step
B) Guiding the model through a logical sequence of intermediate steps, making its reasoning transparent and debuggable
C) Forcing the model to retrieve information from a connected database before answering
D) Eliminating hallucinations by restricting the model to factual statements only

Question 4

When handling API rate limit errors in a production LLM application, what is the recommended retry strategy?

A) Immediately retry the request in a tight loop until it succeeds
B) Switch to a different LLM provider for every request to distribute load
C) Implement exponential backoff, waiting progressively longer between retry attempts
D) Cancel the request and inform the user that LLM APIs are unavailable

Question 5

Few-shot prompting involves providing examples of the desired input-output pattern to the model. What is a key consideration when selecting which examples to include?

A) Always use the maximum number of examples the context window allows to give the model as much information as possible
B) Select examples that are as long as possible so the model learns to produce verbose outputs
C) Choose representative examples that clearly demonstrate the target pattern without overwhelming the context window
D) Use only negative examples (showing what not to do) so the model learns by contrast

Question 6

In a production LLM system, "semantic validation" of model outputs refers to which of the following?

A) Verifying that the model's response is valid JSON or XML
B) Checking that the response makes logical sense within the domain context, not just that it conforms to a structural format
C) Measuring the semantic similarity between the prompt and the response using cosine similarity
D) Running the response through a grammar checker to ensure proper English

Question 7

LLM API costs are typically structured around token consumption. Which of the following strategies most directly reduces the per-request cost of an LLM application?

A) Increasing the temperature to generate shorter responses
B) Caching responses to frequently repeated prompts and implementing prompt engineering to reduce token counts
C) Sending all requests in a single large batch regardless of urgency
D) Disabling response validation to avoid additional API calls

Question 8

In a synoptic tagging workflow for legal document analysis, what role does an LLM play in pattern recognition across documents?

A) It rewrites all documents into a standardized legal format before any analysis occurs
B) It identifies recurring structures and standard clauses while accounting for variations in language and phrasing across documents
C) It replaces human lawyers entirely by rendering legally binding interpretations
D) It performs keyword matching on pre-specified terms to flag documents for review

Question 9

Human-in-the-loop systems in production LLM applications are primarily triggered by which condition?

A) Every request, to ensure a human reviews all AI-generated content before it reaches users
B) Cases where model confidence falls below a threshold, unusual patterns are detected, or specific risk factors are identified
C) Only when the model returns a structured JSON response that fails to parse correctly
D) Requests that exceed the model's context window, requiring the human to summarize the input

Question 10

Batch processing of LLM requests can improve throughput and reduce cost, but it must be implemented carefully. What is the primary tension that makes naive batching problematic?

A) Batching requires all requests in the batch to use identical system prompts
B) Grouping requests together can introduce unacceptable latency for time-sensitive queries and complicates failure handling within a batch
C) The LLM API will generate longer responses when requests are batched, increasing cost
D) Batching prevents the use of caching since each request in a batch is treated as unique