I85PREP
⌘ K
ACP-220 study materials
ACP-220 study resources

ACP-220 practice questions

One question at a time. Build your ACP-220 understanding.

First 50 of 200 questions. Each one lists the choices, the correct answer, and an explanation. The checker at the top of this page saves progress in this browser when it is available.

Continue in the practice tool above to load more questions, 50 at a time.

Question 1AI Model and Product Knowledge

Which statement best describes how a large language model generates text?

  1. It applies a fixed grammar rule set to assemble sentences deterministically.
  2. It predicts the next token from a probability distribution over its vocabulary. Correct answer.
  3. It searches the internet for each word before writing it.
  4. It retrieves the exact answer from a stored database of documents.

Explanation

Language models are trained to model the conditional probability of the next token given prior context, then sample or select from that distribution. Generation is iterative, with each new token appended to the context.

Question 2AI Model and Product Knowledge

What is a token in the context of a language model?

  1. A single neural network layer inside the transformer.
  2. A unit of text, such as a word or word fragment, that the model processes. Correct answer.
  3. A billing period for usage of the model.
  4. A user authentication credential used to call the API.

Explanation

Tokenization splits text into subword units that map to integer IDs in the model vocabulary. Context limits and pricing are typically measured in these units.

Question 3AI Model and Product Knowledge

A team wants the model to answer questions strictly from a set of internal policy documents. Which approach best fits this requirement?

  1. Fine-tuning on unrelated public text to broaden world knowledge.
  2. Retrieval-augmented generation that supplies relevant document excerpts in the prompt. Correct answer.
  3. Removing the system prompt so the model relies on its own judgment.
  4. Increasing the temperature setting to encourage more creative answers.

Explanation

RAG grounds responses in retrieved source passages so the model can cite and follow provided material. This reduces reliance on parametric memory and makes answers auditable.

Question 4AI Model and Product Knowledge

What does the temperature parameter primarily control during sampling?

  1. The size of the model's context window.
  2. The randomness of token selection by scaling the probability distribution. Correct answer.
  3. The maximum number of tokens the model may output.
  4. The number of documents retrieved for grounding.

Explanation

Temperature rescales logits before the softmax, flattening or sharpening the distribution. Lower values make output more deterministic; higher values increase diversity.

Question 5AI Model and Product Knowledge

Which task is an example of zero-shot prompting?

  1. Retrieving similar reviews from a database before classifying.
  2. Providing three labeled reviews before asking the model to classify a new one.
  3. Asking the model to classify a review as positive or negative without providing labeled examples. Correct answer.
  4. Training the model further on a labeled review dataset.

Explanation

Zero-shot prompting gives only the instruction, with no worked examples in the prompt. Few-shot prompting would include labeled demonstrations.

Question 6AI Model and Product Knowledge

A developer needs the model's response to always be parseable as JSON with specific fields. Which technique is most appropriate?

  1. Shortening the system prompt to a single sentence.
  2. Setting temperature to its maximum value.
  3. Adding the phrase 'please be accurate' to the end of the prompt.
  4. Defining a JSON schema and using structured output or tool-calling features where available. Correct answer.

Explanation

Structured output features constrain generation to a schema, making responses machine-parseable. Where unavailable, a detailed schema in the prompt plus validation is the fallback.

Question 7AI Model and Product Knowledge

What is the main purpose of a system prompt?

  1. To store the full history of user messages for later retrieval.
  2. To set persistent instructions and constraints that shape the model's behavior across the conversation. Correct answer.
  3. To encrypt the API request in transit.
  4. To select which GPU runs the inference.

Explanation

The system prompt establishes role, tone, and rules that apply to the whole interaction. It is processed along with user turns but is not a storage or transport mechanism.

Question 8AI Model and Product Knowledge

Which statement about hallucinations is accurate?

  1. They can be eliminated entirely by using a larger model.
  2. They only occur when the temperature is set to zero.
  3. They are caused by exceeding the context window.
  4. They are fluent outputs that are not supported by facts or provided sources. Correct answer.

Explanation

Hallucinations are plausible-sounding but unsupported or incorrect generations. They stem from the probabilistic nature of the model and gaps in grounding, and mitigation relies on retrieval, verification, and citations.

Question 9AI Model and Product Knowledge

What does the context window of a model represent?

  1. The maximum number of tokens the model can consider in a single request. Correct answer.
  2. The number of fine-tuning examples allowed per job.
  3. The number of users allowed to call the API concurrently.
  4. The time limit for a single inference call.

Explanation

The context window bounds the combined input and output tokens the model can attend to at once. Exceeding it requires truncation, summarization, or chunking.

Question 10AI Model and Product Knowledge

A support bot must answer using only the pasted warranty text, even when the model's general knowledge suggests otherwise. Which prompt design best enforces this?

  1. Ask the model to summarize its training data before answering.
  2. Instruct the model to answer solely from the provided text and to state when the text does not cover the question. Correct answer.
  3. Ask the model to be as helpful as possible and answer from any source.
  4. Raise the temperature so the model explores more possibilities.

Explanation

Explicit grounding instructions plus a fallback for missing information constrain the model to the supplied source. This reduces unsupported extrapolation.

Question 11AI Model and Product Knowledge

Which capability is characteristic of multimodal models?

  1. Accepting and reasoning over combinations of text and images in a single request. Correct answer.
  2. Running only on mobile devices without a network connection.
  3. Guaranteeing factual accuracy on every topic.
  4. Replacing the need for any prompt engineering.

Explanation

Multimodal models process more than one input modality, commonly text and images. They still require careful prompting and do not guarantee factual correctness.

Question 12AI Model and Product Knowledge

What is the primary role of an embedding model?

  1. Executing SQL queries against a relational database.
  2. Mapping text into vectors so that semantically similar inputs are close in vector space. Correct answer.
  3. Compressing images losslessly for storage.
  4. Generating long-form prose responses to user questions.

Explanation

Embedding models convert inputs into dense vectors used for similarity search, clustering, and retrieval. They do not generate conversational text.

Question 13AI Model and Product Knowledge

In a retrieval pipeline, what does reranking accomplish?

  1. It reorders an initial candidate set by relevance using a more precise scoring model. Correct answer.
  2. It converts the user query into a SQL statement.
  3. It permanently deletes low-scoring documents from the index.
  4. It trains the embedding model on new documents.

Explanation

Rerankers score query-document pairs more carefully than fast vector search, improving the ordering of the top results. They do not modify the underlying index.

Question 14AI Model and Product Knowledge

Which statement best describes fine-tuning?

  1. Increasing the context window by editing the model configuration.
  2. Further training a pretrained model on task-specific examples to adapt its behavior. Correct answer.
  3. Adding more retrieved documents to the prompt at inference time.
  4. Changing the sampling temperature for a single request.

Explanation

Fine-tuning updates model weights using a targeted dataset so behavior aligns with a specific task or style. It is distinct from prompt-time techniques like retrieval.

Question 15AI Model and Product Knowledge

A team wants the model to adopt a consistent brand voice across many requests without changing weights. Which approach is most direct?

  1. Setting temperature to zero and removing all instructions.
  2. Reducing the max output tokens to a small number.
  3. Encoding the voice guidelines and examples in the system prompt. Correct answer.
  4. Retraining the model from scratch on brand materials.

Explanation

Persistent style guidance in the system prompt shapes tone without retraining. Few-shot examples of the desired voice can strengthen the effect.

Question 16AI Model and Product Knowledge

What is the purpose of a tool or function call in an agentic system?

  1. To increase the model's parameter count at runtime.
  2. To permanently store the model's weights on the server.
  3. To let the model request an external action, such as a lookup, whose result is returned to it. Correct answer.
  4. To bypass authentication for API requests.

Explanation

Tool calling lets the model emit a structured request for an external function; the application executes it and feeds the result back. This extends the model beyond text generation.

Question 17AI Model and Product Knowledge

Which prompt is most likely to produce a concise, well-scoped answer?

  1. Tell me everything about the policy.
  2. Summarize the policy in three bullet points for new employees. Correct answer.
  3. Write about the policy without further constraints.
  4. Discuss the policy in whatever way seems best.

Explanation

Specifying format, audience, and length gives the model clear constraints, which improves relevance and reduces drift. Vague prompts invite unfocused output.

Question 18AI Model and Product Knowledge

What does a model's knowledge cutoff date indicate?

  1. The expiration date of the API key.
  2. The date after which the context window resets.
  3. The point after which events are generally not reflected in the model's pretrained knowledge. Correct answer.
  4. The date the model must be retrained by the customer.

Explanation

The cutoff marks the end of the training data period, so later events may be unknown unless supplied via retrieval or tools. It does not affect credentials or context handling.

Question 19AI Model and Product Knowledge

A developer notices the model ignores an instruction buried in the middle of a very long prompt. Which explanation is most plausible?

  1. The tokenizer deletes all text after the first paragraph.
  2. Long prompts are always rejected by the API.
  3. The model cannot read more than one instruction per request.
  4. Attention can be uneven across long contexts, so important instructions are better placed prominently. Correct answer.

Explanation

Models may attend unevenly across long inputs, and information in the middle can be underweighted. Repeating key instructions near the start or end and trimming irrelevant text mitigates this.

Question 20AI Model and Product Knowledge

Which metric best measures whether a retrieval system surfaces the documents needed to answer a question?

  1. Recall of the relevant documents in the retrieved set. Correct answer.
  2. Number of tokens in the system prompt.
  3. Average response length of the generated answer.
  4. Latency of the embedding model only.

Explanation

Recall measures the share of relevant documents that appear in the retrieved set, which directly affects whether the answer can be grounded. Other listed metrics do not measure retrieval coverage.

Question 21AI Model and Product Knowledge

What is a key trade-off when choosing a smaller model over a larger one for a task?

  1. Guaranteed elimination of hallucinations.
  2. Higher accuracy on all tasks with no downside.
  3. Lower cost and latency versus potentially reduced reasoning quality on complex inputs. Correct answer.
  4. Unlimited context window with no cost.

Explanation

Smaller models typically offer cheaper, faster inference but may underperform on difficult reasoning or nuanced tasks. The choice depends on task complexity and quality requirements.

Question 22AI Model and Product Knowledge

Which practice best reduces prompt injection risk when a model processes untrusted user content?

  1. Separating untrusted content from instructions and validating any actions the model requests. Correct answer.
  2. Trusting any instruction that appears inside the user content.
  3. Disabling all logging to avoid storing malicious text.
  4. Increasing temperature so the model ignores injected text.

Explanation

Treating retrieved or user-supplied text as data rather than instructions, and gating tool actions with validation, limits injection impact. Logging and monitoring also support detection.

Question 23AI Model and Product Knowledge

What does chain-of-thought prompting encourage a model to do?

  1. Retrieve external documents before answering.
  2. Skip reasoning and output only the final answer.
  3. Produce intermediate reasoning steps before arriving at a final answer. Correct answer.
  4. Store its reasoning in the model weights permanently.

Explanation

Chain-of-thought prompting elicits step-by-step reasoning, which can improve performance on multi-step problems. The intermediate steps are generated text, not stored weights.

Question 24AI Model and Product Knowledge

A model returns a confident but outdated price because the price changed last week. Which mitigation addresses the root cause most directly?

  1. Lower the temperature to zero.
  2. Supply the current price through retrieval or a tool lookup at request time. Correct answer.
  3. Increase the max output tokens.
  4. Ask the model to be more confident in its answers.

Explanation

Outdated facts reflect the limits of pretrained knowledge, so providing current data at inference time addresses the cause. Sampling settings and output length do not update facts.

Question 25AI Model and Product Knowledge

Which statement about few-shot examples in a prompt is accurate?

  1. They demonstrate the desired input-output pattern, which helps the model match format and task expectations. Correct answer.
  2. They replace the need for any task instruction.
  3. They guarantee correct answers on unseen inputs.
  4. They permanently update the model's weights.

Explanation

Few-shot examples show the model the expected pattern and format within the prompt. They guide behavior but do not alter weights or guarantee correctness.

Question 26AI Model and Product Knowledge

Which statement best describes what a foundation model is?

  1. A small classifier trained to label one fixed set of categories
  2. A model trained only on one company's private records for a single workflow
  3. A model trained on broad data that can be adapted to many downstream tasks Correct answer.
  4. A rule-based engine that follows hand-written decision trees

Explanation

Foundation models are pretrained on large, broad corpora and then adapted through prompting, retrieval, or fine-tuning to many tasks. This generality is what distinguishes them from narrow single-purpose models.

Question 27AI Model and Product Knowledge

Which description matches a model's parameters?

  1. The documents supplied to the model inside the prompt
  2. The hardware accelerators used to serve the model
  3. The learned weights adjusted during training that shape its outputs Correct answer.
  4. The user-facing settings such as temperature and top-p

Explanation

Parameters are the internal weights the training process tunes to minimize error. They are fixed at inference time and are separate from runtime settings or input content.

Question 28AI Model and Product Knowledge

What does a model's tokenization step do before generation begins?

  1. Translates the input into English before processing
  2. Compresses the prompt to reduce its size
  3. Splits input text into smaller units the model can process Correct answer.
  4. Encrypts the prompt so it cannot be logged

Explanation

Tokenization converts text into subword or word units that map to numeric IDs the model consumes. Token counts also drive context limits and pricing.

Question 29AI Model and Product Knowledge

Which factor most directly explains why two identical prompts can yield different outputs from the same model?

  1. The context window shrinks with each call
  2. The tokenizer changes the vocabulary between calls
  3. Non-zero sampling temperature introduces randomness in token selection Correct answer.
  4. The model retrains itself after every request

Explanation

With a temperature above zero, the model samples from a probability distribution rather than always taking the most likely token, so outputs vary. Setting temperature to zero makes responses largely deterministic.

Question 30AI Model and Product Knowledge

Which statement accurately describes temperature as a generation setting?

  1. Higher values make the model retrieve more documents
  2. Lower values force the model to cite its sources
  3. Lower values enlarge the model's context window
  4. Higher values flatten the distribution and increase output variety Correct answer.

Explanation

Temperature rescales the next-token probability distribution; higher values make unlikely tokens more probable, producing more varied text. It does not affect retrieval, context size, or citation behavior.

Question 31AI Model and Product Knowledge

A developer needs the model to follow a strict output schema so downstream code can parse it. Which approach is most appropriate?

  1. Shorten the system prompt to a single word
  2. Raise the temperature to encourage creative formatting
  3. Remove all examples from the prompt
  4. Constrain generation with structured output or a JSON schema Correct answer.

Explanation

Schema-constrained or structured output modes restrict generation to valid shapes, making parsing reliable. Prompt-only requests for JSON are less dependable than enforced schemas.

Question 32AI Model and Product Knowledge

Which statement about a model's context window is accurate?

  1. It is the number of requests allowed per minute
  2. It bounds the total tokens the model can attend to in one request Correct answer.
  3. It stores the model's weights between sessions
  4. It is the maximum length of a single output token

Explanation

The context window is the combined budget for input and output tokens the model can process in one call. Content beyond it must be truncated, summarized, or retrieved selectively.

Question 33AI Model and Product Knowledge

Which capability distinguishes a multimodal model from a text-only model?

  1. It can accept or produce more than one data modality, such as images and text Correct answer.
  2. It can run without any memory or compute
  3. It requires no prompt to generate output
  4. It always produces shorter answers

Explanation

Multimodal models process or generate multiple modalities, for example combining image and text inputs. This extends the kinds of tasks they can handle beyond pure text.

Question 34AI Model and Product Knowledge

What is the primary function of an embedding model?

  1. Generating long-form prose in response to questions
  2. Enforcing output schemas during decoding
  3. Counting tokens for billing purposes
  4. Mapping text into vectors so semantic similarity can be computed Correct answer.

Explanation

Embedding models convert content into dense vectors positioned so that semantically similar items are close together. These vectors power search, clustering, and retrieval.

Question 35AI Model and Product Knowledge

A team must answer questions over a large internal document set that changes daily. Which architecture best fits?

  1. Raising the temperature so the model explores more answers
  2. Retrieval-augmented generation that fetches current documents at query time Correct answer.
  3. Fine-tuning the model once on a snapshot of the documents
  4. Relying on the model's pretrained knowledge alone

Explanation

Retrieval-augmented generation supplies fresh source content at query time, so updates appear without retraining. Fine-tuning on a snapshot would go stale as documents change.

Question 36AI Model and Product Knowledge

In a retrieval pipeline, what does the reranking stage do?

  1. Generates the final answer from the retrieved passages
  2. Encrypts the retrieved documents before storage
  3. Reorders an initial candidate set by a more precise relevance score Correct answer.
  4. Splits documents into chunks for indexing

Explanation

Reranking takes the candidates returned by fast retrieval and scores them with a more expensive model to improve ordering. It refines relevance rather than generating answers or preparing the index.

Question 37AI Model and Product Knowledge

What is the main effect of chunk size when preparing documents for retrieval?

  1. It determines the model's parameter count
  2. It sets the maximum output tokens for the answer
  3. It trades retrieval precision against the context available in each passage Correct answer.
  4. It controls the temperature used during generation

Explanation

Smaller chunks sharpen matching but may lack surrounding context, while larger chunks carry more context but can dilute relevance. Chunk size is a precision-versus-context trade-off.

Question 38AI Model and Product Knowledge

A team wants the model to adopt a consistent brand voice across many requests without changing model weights. Which approach fits?

  1. Fine-tune the model on a new dataset for each request
  2. Remove the system prompt so the model improvises
  3. Increase temperature until the tone stabilizes
  4. Provide a system prompt with style guidance and representative examples Correct answer.

Explanation

A system prompt with clear style rules and examples steers behavior at inference time without retraining. This is faster and cheaper than fine-tuning for a tone requirement.

Question 39AI Model and Product Knowledge

Which prompt is most likely to yield a concise, well-scoped answer?

  1. Summarize the attached policy in three bullet points for new hires. Correct answer.
  2. Discuss the policy.
  3. What about it?
  4. Tell me everything you know about the policy.

Explanation

Specifying format, length, and audience narrows the response to exactly what is needed. Vague prompts leave scope open and invite long or off-target answers.

Question 40AI Model and Product Knowledge

A developer finds the model ignores an instruction placed in the middle of a very long prompt. Which explanation is most plausible?

  1. The instruction is converted into a tool call automatically
  2. Attention is uneven across long inputs, so mid-context content can be underweighted Correct answer.
  3. The tokenizer removes sentences it considers unimportant
  4. The model deletes the middle tokens to save memory

Explanation

Models do not attend uniformly across long contexts and often weight the beginning and end more heavily. Placing critical instructions near the start or end, or repeating them, mitigates this.

Question 41AI Model and Product Knowledge

Which metric best measures whether retrieval surfaces the documents needed to answer a question?

  1. Average response latency in milliseconds
  2. Recall of the relevant documents in the retrieved set Correct answer.
  3. Number of tokens in the generated answer
  4. Temperature value used during generation

Explanation

Recall measures how many of the truly relevant documents appear in the retrieved set. If relevant documents are missing, no downstream generation step can recover them.

Question 42AI Model and Product Knowledge

A team needs the model to reason through a multi-step arithmetic word problem. Which technique is most likely to improve accuracy?

  1. Removing all context from the prompt
  2. Chain-of-thought prompting that asks for intermediate steps Correct answer.
  3. Requesting the final answer only, with no reasoning
  4. Raising temperature to the maximum value

Explanation

Eliciting intermediate reasoning steps helps the model track multi-step logic before committing to an answer. Answer-only requests give it less room to work through the problem.

Question 43AI Model and Product Knowledge

A model returns a confident but outdated price because the price changed last week. Which mitigation addresses the root cause?

  1. Increase the temperature so the model sounds less certain
  2. Ask the model to apologize for being outdated
  3. Retrieve the current price from a live source and include it in context Correct answer.
  4. Shorten the prompt to reduce token cost

Explanation

The root cause is stale parametric knowledge, so supplying current data through retrieval or a tool call fixes it. Tone adjustments do not change the underlying facts.

Question 44AI Model and Product Knowledge

A model performs well in testing but degrades in production as user inputs drift from the test distribution. Which response best addresses this?

  1. Assume the drift is random and take no action
  2. Lower the context window to reduce exposure to drift
  3. Increase temperature to make outputs more robust
  4. Monitor production inputs and outputs and refresh evaluation data to match real usage Correct answer.

Explanation

Distribution drift means the evaluation no longer reflects real inputs, so ongoing monitoring and updated evaluation sets are needed to detect and correct degradation. Ignoring it or changing sampling does not fix the mismatch.

Question 45AI Model and Product Knowledge

A team must choose between a large general model and a smaller fine-tuned model for a narrow classification task with tight latency limits. Which reasoning is soundest?

  1. Always pick the largest model because size guarantees better accuracy
  2. Pick the model with the highest temperature setting
  3. Choose based on parameter count alone without testing
  4. Benchmark the smaller fine-tuned model on held-out data and compare accuracy against latency targets Correct answer.

Explanation

The decision should rest on measured task accuracy against the latency requirement, since a specialized smaller model can match a large general model on narrow tasks. Parameter count alone is a poor proxy for task performance.

Question 46AI Model and Product Knowledge

Why can a model produce a fluent but fabricated citation?

  1. It deliberately hides the correct source from the user
  2. Its context window forces it to invent citations
  3. Its tokenizer inserts fake references automatically
  4. It predicts plausible token sequences rather than verifying facts against a source Correct answer.

Explanation

Generation optimizes for plausible continuations, not factual verification, so a citation can be well-formed yet nonexistent. Grounding answers in retrieved sources and requiring verifiable references reduces this risk.

Question 47AI Model and Product Knowledge

A product team wants the model to reliably extract dates from invoices in a fixed format. Which approach most directly constrains the output to that format?

  1. Increasing the maximum token limit
  2. Defining a structured output schema the response must conform to Correct answer.
  3. Adding more examples of free-form prose to the prompt
  4. Raising the temperature setting

Explanation

A structured output schema (for example, a JSON schema) constrains generation so the response conforms to defined fields and types. This makes downstream parsing reliable. Other sampling or length settings do not enforce structure.

Question 48AI Model and Product Knowledge

A team is deciding between fine-tuning and retrieval-augmented generation for giving a model access to a knowledge base that is updated hourly. Which consideration most favors retrieval-augmented generation?

  1. Retrieval eliminates the need for any prompt engineering
  2. Retrieval keeps the knowledge external so updates take effect without retraining Correct answer.
  3. Retrieval guarantees the model will never hallucinate
  4. Retrieval permanently writes new facts into the model's weights

Explanation

In retrieval-augmented generation the knowledge lives in an external store that is queried at inference time, so content changes propagate immediately. Fine-tuning bakes knowledge into weights and would require retraining to reflect frequent updates. Retrieval reduces but does not eliminate hallucination.

Question 49AI Model and Product Knowledge

A developer notices that a model answers questions about recent events incorrectly even when the correct facts appear in the prompt. The facts are placed after a long block of unrelated text. Which explanation best fits this behavior?

  1. The relevant facts fall outside the effective attention of the long context Correct answer.
  2. The temperature was set to zero
  3. The model's weights were updated after the prompt was sent
  4. The tokenizer merged the facts into a single token

Explanation

Information positioned deep within a long context can receive less effective attention, so relevant facts may be underused even when present. Placing key facts near the start or end of the prompt, or trimming irrelevant text, mitigates this. Weight updates and tokenization do not explain it.

Question 50AI Model and Product Knowledge

A developer is designing a system and wants to control how much randomness the model applies when choosing among likely next tokens during generation. Which setting should the developer adjust?

  1. The temperature parameter Correct answer.
  2. The embedding dimensions
  3. The context window size
  4. The maximum output token limit

Explanation

Temperature scales the probability distribution over candidate next tokens, so lower values make output more deterministic and higher values make it more varied. The other settings govern output length, vector representation size, or how much input the model can accept, none of which control sampling randomness.