Free AIP-C01 practice questions

Choosing the lightest technique that meets the requirement: prompt engineering before retrieval or model customization.

The model's knowledge is already adequate; the gap is output structure. Structure is controlled most cheaply by the prompt: explicit headings and an example response. Fine-tuning also works but needs labeled pairs and a training job, continued pre-training changes knowledge rather than format, and retrieval exists to inject information the model does not have, not to impose a fixed layout.

1. Name the gap: knowledge, freshness, or behaviour/format.
2. For a format gap with a capable model, try prompt instructions and examples first.
3. Reserve fine-tuning for when prompting cannot hold the behaviour and labeled data exists.
4. Use retrieval only when the model lacks information.

Exam tip: Fix format with the prompt first; customize the model only when prompting cannot hold the behaviour.

Strands separates persistence (session managers) from context-window management (conversation managers).

The failure is a lost session across horizontally scaled tasks, so state must be stored somewhere every task can read. A session manager backed by S3 saves the messages and agent state per session ID and restores them on any task. A file-system session manager only works when every request hits the same host, and both conversation managers shape the in-memory context without persisting it.

1. Notice that requests for one conversation reach different containers.
2. Recognise that the agent's state must live outside the container.
3. Pick the Strands component that persists sessions: a session manager with shared storage.
4. Reject conversation managers, which only trim or summarise context in memory.

Exam tip: Session managers persist an agent's state; conversation managers only manage what fits in the context window.

Amazon Bedrock Guardrails content filters (hate, insults, sexual, violence, misconduct, prompt attack) can be applied to user prompts and model responses with configurable strength.

A guardrail passed in guardrailConfig evaluates the input first; if a filter blocks it, the model is not invoked and the blocked-prompt message is returned (stopReason guardrail_intervened). It is managed configuration, so it beats a Comprehend-plus-Lambda pipeline or a custom classifier on overhead, and unlike a system prompt it does not depend on the model choosing to comply.

1. Identify the need: stop harmful input before the model, with a fixed reply.
2. Note the constraint: no classification logic to build or host.
3. Match it to a managed input control on the Bedrock call itself.
4. Reject options that add code, training, or rely on the model's cooperation.

Exam tip: Harmful-input blocking with a canned reply and no custom code = a Bedrock guardrail content filter on prompts.

Token tracking and cost attribution: Bedrock costs can be split with tagged application inference profiles, and individual requests can be found through invocation logs filtered on requestMetadata.

Two different needs are stated. Spend in Cost Explorer needs a billable resource that carries a cost allocation tag: an application inference profile per tenant, used as the modelId, does that without new roles or accounts. Request-level evidence needs the full prompt and response: model invocation logging captures them, and requestMetadata (up to 16 key-value pairs per request) adds the tenant ID so the logs can be filtered. CloudWatch token metrics are not spend, the shared Lambda tag cannot separate tenants, and CloudTrail holds neither tenant nor content.

1. Split the requirements: billing attribution and request-level forensics.
2. For billing, find a Bedrock resource that can carry a per-tenant cost allocation tag.
3. For forensics, find a log that stores prompts and responses and can be filtered per tenant.
4. Reject options that give token counts, shared tags or API-call records without content.

Exam tip: Per-tenant Bedrock spend: tagged application inference profiles. Per-tenant request lookup: invocation logs plus requestMetadata.

Open-ended FM output can be correct in many wordings, so lexical metrics (exact match, n-gram overlap) under-score it. Semantic metrics compare meaning.

The failures are paraphrases, so the score must compare meaning rather than tokens. BERTScore embeds both texts with a transformer and matches them token by token in embedding space, which credits synonyms and reworded sentences. ROUGE and exact match both depend on shared surface words, and perplexity measures fluency-like predictability with no reference at all.

1. Note the symptom: correct answers fail because the wording differs.
2. Rule out metrics that compare surface tokens (exact match, ROUGE).
3. Rule out reference-free fluency signals such as perplexity.
4. Choose an embedding-based similarity to the reference.

Exam tip: Paraphrase-tolerant accuracy needs a semantic metric such as BERTScore, not word overlap.