Free AIF-C01 flashcards

When deciding whether a single foundation model call can process an entire 200,000-token source document without retrieval augmentation, which model property is the deciding constraint?

The context window. It is the maximum number of tokens a model can process per inference call, covering both input and output. If the source exceeds this window, the application must use RAG or chunking to surface relevant content first.

A model's training data cutoff date is what limits whether it can process a large document supplied at inference time.

Confuses baked-in knowledge with input capacity. Training cutoff affects facts stored in the weights, not the ability to process tokens supplied at inference. Providing the document as context bypasses the cutoff entirely; the limiting factor for input is the context window.

A hosted LLM returns noticeably different wording each run for the same prompt. Which sampling change keeps answers closely on-message with far less variation, without forcing identical output?

Lower the temperature toward 0 (for example 0.2). Temperature scales how sharply the model favours its most-probable tokens; lowering it concentrates probability mass on the top tokens, so the same prompt yields closely similar, on-message wording each time while still allowing slight variation.

To make an LLM's repeated answers more consistent, raise the temperature toward 1.0 so it considers a wider spread of next tokens.

Wrong direction. Higher temperature flattens the token distribution and increases variation, moving away from on-message consistency. To reduce variation you lower temperature, which sharpens the distribution toward the likeliest tokens.

Which AI discipline lets a recommendation system improve its accuracy over time from customer purchase history, without engineers coding new rules each season?

Machine learning. ML algorithms discover patterns in training data and adjust their internal parameters so predictions improve as more data is seen. This self-improving property from historical data is what separates ML from rule-based systems that need a human to codify every new condition.

A rule-based expert system that encodes domain knowledge as hand-crafted if-then rules can improve its recommendations autonomously as new data arrives.

Category error. Rule-based systems require explicit programming of every new rule and do not learn autonomously from data. The ability to improve from experience without re-programming is the defining property of machine learning, not of rule-based expert systems.

How does few-shot prompting improve a classifier's accuracy on edge cases without any fine-tuning?

Add labelled input-output pairs for each category directly in the prompt. These examples demonstrate the desired mapping, narrowing the model's interpretation of the task and anchoring its output to the correct label space, which reduces wrong labels for inputs near category boundaries.

Raising the temperature makes a model classify edge-case inputs more accurately because it explores more options.

Wrong lever. Higher temperature increases randomness, which worsens classification consistency rather than improving it. It also does not address the root cause of misclassification - the model's ambiguity about which label to assign.