cert-bank

Guides

이 글을 한국어로 읽기

AI-103 Study Guide: Follow One Request Through a Foundry App, Hop by Hop

AI-103 is a developer exam about building generative AI apps and agents on Microsoft Foundry, and it is full of short code and request settings. The easiest way to hold it together is to follow a single user request through an app: it is authenticated, sent to a deployment, grounded with your data, possibly handed to an agent and its tools, checked by guardrails, and finally logged. Each hop has its own decisions and its own typical failure. This guide walks those hops in order.

Study guide · Microsoft Azure AI Apps and Agents Developer Associate (AI-103)

AI-103 Study Guide: Follow One Request Through a Foundry App, Hop by Hop

The exam

120 minutes, passing score 700 on a scale of 1 to 1,000. Microsoft does not publish a fixed number of questions, so at the halfway mark check that you are about halfway through the questions. Skill areas: Plan and manage an Azure AI solution (25–30%), Implement generative AI and agentic solutions (30–35%), and computer vision, text analysis and information extraction (10–15% each). More than half the exam is planning plus generative AI and agents.

Among the 243 AI-103 questions we wrote for this question bank, almost a third (32%) are troubleshooting questions: a call returns 401, 404 or 429, an agent ignores a tool, citations are missing. Microsoft Foundry itself appears in 40% of them, followed by Azure AI Search and Foundry Agent Service (16% each) and Azure OpenAI (15%). Expect Python and REST snippets.

AI-103: follow one request through the app, hop by hop
1
Identity and endpoint
Entra ID over keys; v1 endpoint for the OpenAI client, project endpoint for the Foundry SDK
2
Deployment and quota
Model takes the deployment name; 429s count prompt tokens plus max_tokens
3
The call
Responses (stateful) or Chat Completions (stateless); most reasoning models drop temperature
4
Grounding
Hybrid query with semantic ranking; security trimming in the query filter, not the prompt
5
Agent and tools
Deployment plus instructions plus tools; your app runs function calls, not Foundry
6
Guardrails, then logging
Guardrails enforce, a system message only steers; traces need Application Insights
Each card is one hop of a single request, in the order this guide covers them; each has its own typical failure.

Hop 1: who is calling, and at which endpoint

Two endpoints exist and they are not interchangeable. The resource's v1 endpoint (https://<resource>.openai.azure.com/openai/v1/) is used with the standard OpenAI client. The project endpoint (https://<resource>.services.ai.azure.com/api/projects/<project>) is used by the Foundry SDK: AIProjectClient handles project operations such as listing connections, and its OpenAI-compatible client handles responses, agents and evaluations. Give the project client the v1 address and listing connections fails with 404.

For identity, Microsoft recommends Microsoft Entra ID over keys: DefaultAzureCredential (your az login locally, a managed identity in Azure) and the Foundry User role on the project. Keyless code alone does not switch keys off; disabling local authentication does.

Going deeper: the 401 that appears only after an hour

get_bearer_token_provider(DefaultAzureCredential(), "https://ai.azure.com/.default") returns a function, not a token. Pass the function as api_key and the client calls it whenever it needs a fresh token. Call it once and pass the resulting string, and everything works in testing, then every call fails with 401 once that token expires, until the process restarts. Questions that describe "worked this morning, 401 since lunch, no code change" are pointing at exactly this.

Hop 2: which deployment, and is there room

The model parameter takes your deployment name. The deployment type decides where prompts are processed (Global, Data Zone, or the resource's own geography) and how you pay (per token, provisioned throughput units, or Batch with a 24-hour target). Quota is tokens per minute, per model, region and deployment type, and each deployment takes a slice of it.

The classic 429 trap: the rate limiter counts an estimate made on arrival, prompt tokens plus max_tokens, not the tokens you are billed for. An oversized max_tokens causes 429s while usage graphs look low. A 429 with unused regional quota often means the deployment itself is under-allocated.

Hop 3: the call itself

  • Responses or Chat Completions. Responses is the newer, stateful API: continue a conversation with previous_response_id. Chat Completions is stateless: your app resends the message list. Structured outputs go in response_format for Chat Completions and in text.format for Responses.
  • Reasoning models. Most of them do not support temperature and top_p, so drop them; use max_completion_tokens or max_output_tokens and tune reasoning_effort. The budget includes hidden reasoning tokens, so a small budget returns an incomplete answer.
  • Images in the prompt. The image URL must be reachable by the service; for a private file, send base64 data instead.

Hop 4: grounding with your data

Classic RAG: embed the question with the same embedding model used for the index, run a hybrid query with semantic ranking, keep the top chunks, label each with an id, and tell the model to answer only from them and cite the ids. Three rules carry many questions:

  • Security trimming belongs in the query (an OData filter on the user's groups), not in the instructions. The model sees whatever you retrieve.
  • Hybrid scores are Reciprocal Rank Fusion scores and stay tiny (around 0.03 at best for one text plus one vector query), so a cut-off tuned for cosine similarity, such as 0.8, removes every result. Semantic reranker scores run from 0 to 4.
  • Azure OpenAI On Your Data (the data_sources block) retired on 14 October 2026. Removing the block is not a migration; something must still retrieve the documents.

Beginner tip: run one hybrid query against an Azure AI Search index from the portal's Search explorer and look at the @search.score values, then add a semantic configuration and look at @search.rerankerScore. Seeing scores of 0.03 next to scores of 2.7 for the same documents makes the "why did my threshold remove everything?" questions obvious.

Hop 5: the agent and its tools

A prompt agent in Foundry Agent Service is a deployment plus instructions plus tools, saved as immutable versions. A conversation holds the history; a response is one run. Know which knowledge tool fits: file search for documents a user brings to one conversation, the Azure AI Search tool for an existing index, a Foundry IQ knowledge base (some features still in preview) for curated knowledge shared by many agents. With function calling, Foundry never runs your code: the model returns a function call, your app runs it and sends back the output. If the agent skips a tool you need, tighten the instructions or set tool_choice to required.

For several agents, know the orchestration patterns in Microsoft Agent Framework: sequential for dependent steps, concurrent for independent views in parallel, handoff to pass control at run time. Foundry portal workflows retire on 1 December 2026; Agent Framework is the code path.

Hop 6: guardrails

Guardrails (previously content filters) enforce safety on inputs and outputs; a system message only steers. Harm categories (hate, sexual, violence, self-harm) are scored by severity; Prompt Shields detect jailbreaks and indirect (document) attacks in text, and document attack detection is off until you enable it; blocklists match terms. Instructions written inside an image must first be extracted as text before Prompt Shields can check them. When the stem says the model "must never" do something, the answer is a control that enforces, not a line in the prompt.

Hop 7: what you see afterwards

Azure Monitor gives platform metrics (requests by status code, time to first token, blocked volume). Agent traces need an Application Insights resource connected to the project; without it the tracing views stay empty. Evaluation results are stored in the project and can also be sent to Application Insights. Pin model versions and compare evaluation scores before you move to a new one, because drift often follows a version upgrade.

Check the status of every feature you pick

FeatureStatus in the bank's sources
gpt-image-2Generally available
dall-e-3Retired
Sora 2, agent MemoryPreview: not a key answer for a production need
Azure OpenAI On Your DataRetired 14 October 2026
Classic threads, messages and runsRetiring 31 March 2027, replaced by conversations and responses

Microsoft says most questions cover generally available features. When two answers both work and one relies on a preview or retired feature, choose the other.

Practise

Practise in the question bank: every question explains each option, right and wrong. Start with the domain you just studied, then take a timed mock exam.

Skills measured as listed in Microsoft's AI-103 study guide. Foundry features, names and retirement dates change often; check the exam page and Microsoft Learn for the current status before your exam.