AI-103 exam questions explained by domain (with answers)

Question 1. An order agent must retrieve policy documents and query a live transactional API. How should the two capabilities be represented? Options: a knowledge source for policies and a tool for the API; both as permanent model weights; both only as conversation history; a content filter for the API. Answer: knowledge source plus tool. Knowledge sources ground informational answers; tools perform actions and obtain current transactional data. This is what the outline calls choosing "memory, tool, and knowledge integration services for agent solutions".

Question 2. A support app classifies short requests on edge devices with strict latency and cost limits. Which model type should be evaluated first? Answer: a small language model optimized for the task. The largest multimodal model, an image generator or a speech model add capabilities the scenario does not need and break its constraints. The pattern repeats across the domain: the stem fixes two or three constraints and only one option satisfies all of them.

Question 3. A RAG app retrieves passages, but the model ignores source identifiers. What should context assembly include? Options: explicit source labels and instructions to cite supplied evidence; higher temperature without sources; the entire index without ranking; a random citation after generation. Answer: source labels and citation instructions. They tie each claim to a retrieved passage; post-processing cannot invent valid citations and a higher temperature makes things worse.

This is the heaviest domain and the one that changes most from AI-102: deploying and consuming models, RAG, tool-augmented flows, evaluating fabrications and relevance, agents with memory and function calling, orchestrated multi-agent solutions, human approvals, tracing and token analytics. Questions usually give a symptom (answers without citations, high latency, an agent calling the wrong tool) and ask for the smallest change that fixes it.

Question 4. A workload performs one predefined extraction task at high volume; another requires richer multi-step analysis. How should Content Understanding modes be selected? Answer: single-task for focused extraction and pro mode for complex analysis. The choice reflects complexity, throughput, latency and required reasoning; using pro mode for everything costs more and runs slower for no reason.

Question 5. A RAG corpus contains PDFs, scans, images and recorded interviews. What ingestion design is appropriate? Answer: route each modality through suitable extraction (OCR, layout analysis, transcription) before chunking and indexing normalized content, keeping the metadata. Treating every binary as plain text, indexing filenames only or discarding metadata leaves an index you cannot retrieve from. The text and speech domain, not shown here, asks in a similar way about entity extraction, translation and speech as an agent modality.

Each describes a system with concrete constraints and offers one correct option against distractors that fail one of them. What works is underlining the constraints in the stem (live, edge, cite, high volume, several modalities) and discarding each option by asking which one it breaks. If you are torn between two, one almost always solves the problem and the other solves a neighbouring problem.

The AI-103 practice test has hundreds of questions with this structure, an online test by domain to see where you fail and a timed 120-minute mock exam. To organise your study, see the AI-103 outline breakdown and the eight-week plan.