Free AI-103 practice test: 30 questions with answers

Function calling exposes a named tool with a typed parameter schema, and the model returns structured arguments that the application then executes. Pasting documentation into the prompt or fine-tuning does not give the runtime a reliable, callable interface.

Business rules belong in deterministic code, because a prompt instruction is a suggestion the model can ignore. A tool that refuses over-limit refunds and returns an approval status keeps the control outside the model.

Narrow roles and tool sets reduce what each agent can get wrong, and an orchestrator decides which specialist handles each step. Agent handoffs do not remove the need for evaluation, and model outputs remain non-deterministic.

A per-turn budget bounds the loop in code rather than relying on the model's judgment. Removing the tool loses capability, and raising a rate limit lets the runaway loop spend more.

Hybrid search runs a full-text query and a vector query together and merges the ranked results, so exact codes and paraphrased questions both match. Facets and suggesters do not perform semantic retrieval, and a wildcard filter does not rank by meaning.

The retrieved chunks are the problem, so the fix belongs in indexing: splitting on structure keeps each table and section intact. Temperature and model size do not repair missing context, and returning one result usually removes evidence the model needs.

Grounding depends on which evidence reaches the model, which is set by retrieval count and relevance, and on how well the source was chunked at indexing time. Output limits, UI theme, and log retention do not change what the model is able to cite.

An embedding model maps text to vectors where nearby points have similar meaning, which is what vector search compares. Tokenizers, rerankers, and filters do not produce the stored vectors used for retrieval.

A smaller model answers faster and costs less per call, which suits a narrow classification task. Large reasoning models add latency, and a nightly batch cannot serve interactive requests.

A maximum output token setting stops generation at a fixed length, which bounds both latency and cost. Logit bias and embedding size do not limit output length, and a longer system prompt increases input cost.

A canary release sends a fraction of traffic to the new version and lets the team compare results before expanding. A big bang cutover or deleting the old deployment first gives no controlled comparison.

Exponential backoff with jitter spreads retries out and respects the service's guidance, which avoids making throttling worse. Tight loops amplify load, and changing identity to dodge quotas is not a valid remedy.

Groundedness measures whether each claim in the answer can be traced to the retrieved passages. Perplexity measures language-model fluency, and throughput and token counts measure performance and cost.

A batch evaluation on the same dataset gives comparable scores for both versions, which shows whether the change helps. A single manual chat or a code review does not measure quality across representative inputs.

Running evaluation automatically in CI catches regressions on every change, before release. Manual quarterly reviews and user complaints are slow, and formatting checks say nothing about answer quality.

Groundedness measures support from context and relevance measures whether the answer addresses the question. High groundedness with low relevance means the answer is faithful to retrieved text but off-topic, often because retrieval fetched the wrong passages or the prompt is unfocused.

A managed identity lets the app obtain Microsoft Entra ID tokens without storing a secret, and a role assignment limits what it can do. Hard-coded or shared keys leak easily and cannot be scoped to a single workload.

Key Vault stores secrets centrally, supports rotation, and lets the app pick up a new version without a code change or redeployment. Config files in source control force a redeploy and expose the value to everyone with repository access.

Retrieval at query time reads the latest indexed content, and a scheduled indexer keeps it fresh. Fine-tuning is slow and costly for changing facts, and a build-time copy is stale between deployments.

Azure AI Content Safety analyzes text and images for harmful categories and can block content before it is returned. Monitoring, edge firewalls, and backup do not evaluate the meaning of generated text.

Any script shipped to the browser is readable by every visitor, so the key is effectively public. TLS protects data in transit but not a secret that is delivered to the client, so a backend proxy should make the model call.

Disclosure makes the system's nature clear to users, and avoiding decisions based on protected attributes addresses fairness. Scalability, availability, and cost are operational concerns rather than responsible AI principles.

Application Insights stores distributed traces, and OpenTelemetry instrumentation lets the agent emit spans for each model and tool call. The other services do not record per-request execution traces.

Token usage drives cost, latency percentiles show user experience, and 429 responses reveal quota pressure. License counts and UI colors do not describe the health or cost of the model endpoint.

Linking feedback to traces lets the team inspect the exact run that failed and turn it into a regression test. Deleting logs, ignoring ratings, or watching only CPU does not reveal answer quality problems.

Delegated, scoped identity ensures the downstream API applies the user's permissions, and server-side authorization enforces access regardless of what the model requests. Shared admin credentials and disabled authentication remove the boundary entirely.

Key Vault with managed identity keeps secrets out of code, and scheduled or emergency rotation limits the impact of exposure. Frontend bundles, shared keys, and logging authorization headers each expose credentials.

Entra ID sign-in with a scoped role assignment gives each developer audited, revocable access without a shared secret. Emailing or sharing keys and disabling authentication break the stated policy.

Least privilege limits what a hijacked agent can do, and separating retrieved content from instructions reduces the chance that page text is obeyed as a command. Temperature and logging do not stop injected instructions, and copying the system prompt into documents leaks it.

Destructive actions need an explicit approval gate, and the tool's credentials should only permit what it needs. Prompt examples and temperature changes reduce error rates only slightly and do not stop a wrong call from executing.