AI-200 · Domain 2 of 4

AI-200 Domain 2: Develop AI solutions by using Azure data management services

This is the largest and most distinctive AI-200 domain. It covers the data layer of AI apps: vector search and queries in Azure Cosmos DB for NoSQL, pgvector and RAG patterns in Azure Database for PostgreSQL, and caching and vector indexing in Azure Managed Redis.

Exam weight
25–30%
Official functional groups
3
Estimated questions estimate
11–16
Suggested study time estimate
20–30 hrs

About 11 min read Reviewed October 2026

Overview

Develop AI solutions by using Azure data management services is the largest domain and the one that most distinguishes AI-200 from the retired AZ-204. It has three objective groups. Cosmos DB for NoSQL covers connecting with the SDK and running queries, optimizing query performance and Request Units with indexing and consistency levels, storing and retrieving embeddings and running vector similarity search for semantic retrieval, and the change feed. PostgreSQL covers connecting and querying with SDKs, schema and indexing strategies, configuring compute and storage for vector workloads, running vector similarity search with pgvector including retrieval-augmented generation (RAG) with metadata filtering, and connection optimization. Azure Managed Redis covers caching, expiration and invalidation, and vector indexing for similarity search.

At 25–30%, this is where the exam concentrates, and vector search is the single biggest new area. The exam rewards knowing how to store embeddings, run similarity search, and build RAG patterns across these data services.

How it's tested

AI-200 questions in this group describe an AI data or retrieval need and ask which service and approach fits. AI-200 questions often describe a requirement without naming the service, so read the requirement and make the service choice defensible from it before answering.

  • Vector search choice

    A semantic-retrieval need is described, and you choose vector search in Cosmos DB, PostgreSQL pgvector, or Redis.

  • Optimize the data path

    A performance or cost concern is described, and you choose indexing, consistency level, RU optimization or connection tuning.

  • Build RAG

    A retrieval-augmented generation need is described, and you choose the embedding storage and metadata-filtered similarity search.

Reasoning it through

An application must retrieve the most semantically similar documents to a user query to ground a language model's answer, using a relational store the team already runs on PostgreSQL. Full-text search returns keyword matches, not semantic similarity. Storing embeddings in Azure Database for PostgreSQL with pgvector and running a vector similarity search, filtered by metadata, retrieves the closest documents for a RAG pattern, which is the approach the exam expects for semantic retrieval on PostgreSQL.

Functional groups 3

Official functional groups from the Microsoft outline, in outline order.

2.1 Develop AI solutions by using Azure Cosmos DB for NoSQLSDK queries, RU and consistency optimization, embeddings, vector search and change feed.
What it means
This covers connecting to Azure Cosmos DB for NoSQL with the SDK and running queries, optimizing query performance and Request Unit (RU) consumption with indexing policies and consistency levels, storing and retrieving embeddings and running vector similarity search for semantic retrieval, and implementing a change feed processor to handle new or updated items.
What to understand
Cosmos DB charges in Request Units, so indexing policies and the right consistency level control cost and performance. It can store embeddings and run vector similarity search for semantic retrieval. The change feed streams item changes for reactive processing. Choosing the consistency level trades latency against guarantees.
How it shows up
A Cosmos DB performance, cost or vector-search need is described, and you must choose indexing, consistency, vector search or the change feed.
Common confusion
Confusing consistency levels; stronger consistency increases latency and RU cost, while weaker consistency is cheaper and faster but less strict.
2.2 Develop AI solutions by using Azure Database for PostgreSQLSDK queries, schema and indexing, vector workloads, pgvector similarity search and RAG.
What it means
This covers connecting and querying Azure Database for PostgreSQL with SDKs, modeling schemas and indexing strategies with appropriate data types, optimizing query latency and reducing pgvector compute overhead, configuring compute, memory and storage for vector workloads, running vector similarity search with pgvector including storing embeddings and implementing retrieval-augmented generation (RAG) with metadata filtering, and optimizing connections for throughput and latency.
What to understand
PostgreSQL with the pgvector extension stores embeddings and runs vector similarity search, supporting RAG where retrieved context grounds a model's response. Indexing and resource configuration control vector query latency and cost. Metadata filtering narrows results. Connection pooling optimizes throughput.
How it shows up
A PostgreSQL vector, RAG or performance need is described, and you must choose pgvector search, indexing, resource configuration or connection optimization.
Common confusion
Confusing vector similarity search with full-text search; vector search finds semantically similar items by embeddings, while full-text search matches keywords.
2.3 Integrate Azure Managed Redis in AI solutionsCaching, expiration and invalidation, and vector indexing.
What it means
This covers Azure Managed Redis data operations including caching, expiration and invalidation, and implementing vector indexing to enable similarity search in Redis.
What to understand
Redis caches frequently accessed data in memory to cut latency and cost, with expiration and invalidation keeping the cache fresh. It can also index vectors for fast similarity search, serving as a low-latency vector store. Caching embeddings or results reduces repeated expensive computation.
How it shows up
A caching or low-latency vector need is described, and you must choose Redis caching, expiration, invalidation or vector indexing.
Common confusion
Treating Redis only as a cache; it can also serve as a fast vector index for similarity search.

High-yield concepts

The terms, rules and formulas this domain turns into questions most often.

  • Embeddings

    Numeric vector representations of data for semantic comparison.

    Remember
    Similar meaning produces nearby vectors.
    Tested as
    Vector search questions.
    Trap
    Confusing embeddings with keywords.

  • Vector similarity search

    Finding the nearest vectors to a query embedding.

    Remember
    It retrieves semantically similar items.
    Tested as
    Retrieval questions.
    Trap
    Confusing it with full-text search.

  • Retrieval-augmented generation (RAG)

    Grounding a model with retrieved relevant context.

    Remember
    Retrieve, then generate from the context.
    Tested as
    RAG questions.
    Trap
    Generating without retrieval grounding.

  • Cosmos DB Request Units

    The throughput and cost currency of Cosmos DB.

    Remember
    Indexing and queries affect RU cost.
    Tested as
    Cost questions.
    Trap
    Ignoring RU consumption.

  • Consistency levels

    Trade-offs from strong to eventual consistency.

    Remember
    Stronger consistency costs more latency and RUs.
    Tested as
    Consistency questions.
    Trap
    Choosing strong where eventual suffices.

  • Change feed

    A stream of item changes in Cosmos DB.

    Remember
    It enables reactive processing of changes.
    Tested as
    Event processing questions.
    Trap
    Polling instead of using the change feed.

  • pgvector

    PostgreSQL extension for storing and searching vectors.

    Remember
    It enables vector similarity search in PostgreSQL.
    Tested as
    PostgreSQL vector questions.
    Trap
    Expecting vectors without the extension.

  • Metadata filtering

    Narrowing vector search by attributes.

    Remember
    It combines similarity with filters.
    Tested as
    RAG questions.
    Trap
    Returning irrelevant matches without filters.

  • Redis caching

    In-memory caching with expiration and invalidation.

    Remember
    It cuts latency and repeated computation.
    Tested as
    Caching questions.
    Trap
    Serving stale data without invalidation.

  • Redis vector indexing

    Indexing vectors in Redis for similarity search.

    Remember
    Redis can be a low-latency vector store.
    Tested as
    Vector questions.
    Trap
    Treating Redis as cache only.

  • Connection optimization

    Pooling and tuning connections for throughput.

    Remember
    It minimizes latency at scale.
    Tested as
    Performance questions.
    Trap
    Opening a connection per request.

Side by side

Concepts candidates mix up, compared on the point that decides the answer.

Vector search vs. full-text search

ConceptMeaningKey differenceCommon trap
Vector similarity search Finds semantically similar items by embeddings Meaning-based retrieval Confusing with keyword search
Full-text search Matches keywords and terms Literal text matching Expecting semantic results

Cosmos DB vs. PostgreSQL pgvector vs. Redis for vectors

ConceptMeaningKey differenceCommon trap
Cosmos DB for NoSQL NoSQL with vector search and change feed Global scale, RU-based Ignoring RU cost
PostgreSQL pgvector Relational with pgvector and RAG Rich SQL and metadata filtering Expecting vectors without pgvector
Azure Managed Redis In-memory cache and vector index Lowest latency Treating it as cache only

Cosmos DB consistency levels

ConceptMeaningKey differenceCommon trap
Strong Latest data always Highest latency and RU cost Using it where eventual suffices
Session / bounded Balanced guarantees Common default trade-off Misjudging the guarantee
Eventual Fastest, least strict Lowest latency and cost Using it where strong is required

Exam traps

Each trap is an answer that looks right. Learn the clue that should change your mind.

  1. Full-text for semantic

    What candidates pick
    Use full-text search to find semantically similar documents.
    Why it looks right
    It searches text.
    Why it's wrong
    Full-text matches keywords; vector similarity search on embeddings finds semantically similar items.

    Clue to watch for The retrieval must be by meaning, not exact words.

  2. Strong consistency always

    What candidates pick
    Choose strong consistency for every Cosmos DB operation.
    Why it looks right
    It guarantees the latest data.
    Why it's wrong
    Strong consistency adds latency and RU cost; weaker levels suffice for many AI retrieval workloads.

    Clue to watch for The workload tolerates slightly stale reads and cost matters.

  3. RAG without retrieval

    What candidates pick
    Call the model directly without retrieving relevant context.
    Why it looks right
    It is simpler.
    Why it's wrong
    RAG retrieves relevant context first and grounds the generation on it, improving accuracy.

    Clue to watch for The answer must be grounded in specific data.

  4. Vectors without pgvector

    What candidates pick
    Expect PostgreSQL to run vector similarity search without the extension.
    Why it looks right
    PostgreSQL is powerful.
    Why it's wrong
    Vector storage and similarity search in PostgreSQL require the pgvector extension.

    Clue to watch for The task is vector search on PostgreSQL.

  5. Redis is cache only

    What candidates pick
    Rule out Redis for a low-latency vector store because it is a cache.
    Why it looks right
    Redis is known as a cache.
    Why it's wrong
    Azure Managed Redis also supports vector indexing for similarity search as a low-latency vector store.

    Clue to watch for A very low-latency vector search is needed.

  6. Connection per request

    What candidates pick
    Open a new database connection for every request.
    Why it looks right
    It is straightforward.
    Why it's wrong
    Connection pooling and optimization improve throughput and reduce latency at scale.

    Clue to watch for The workload is high-throughput.

Answer framework

For any Domain 2 scenario, use these steps to find the answer AI-200 expects.

  1. Identify the data need: NoSQL scale, relational and RAG, or in-memory low latency, and whether vector search is required.
  2. For vector similarity search, store embeddings and query by similarity, choosing Cosmos DB, PostgreSQL pgvector, or Redis to match scale, SQL needs and latency.
  3. For RAG, retrieve relevant context with metadata-filtered similarity search, then ground the generation on it.
  4. Optimize the data path, using Cosmos DB indexing and the right consistency level, pgvector indexing and resource sizing, and connection pooling.
  5. Use Redis for caching with expiration and invalidation, and for fast vector indexing where lowest latency is required. If filters on relational metadata matter, prefer PostgreSQL with pgvector.

Practice question

A developer must build a retrieval-augmented generation feature that finds the most semantically similar documents to a user's question to ground a language model's response, using the team's existing Azure Database for PostgreSQL. Which approach is MOST appropriate?

Want more like this? The AI-200 practice test has questions for every domain.

Checklist

0 of 10

Tick each one you can do without notes. Your ticks stay in this browser.

Study strategy

Give this domain the most study time, because it is the largest and the furthest from traditional Azure development. If vector search is new to you, start there: understand embeddings, vector similarity search, and the RAG pattern conceptually, then practice each across the three data services.

Learn how each service handles vectors and when to choose it: Cosmos DB for NoSQL scale with RU and consistency trade-offs, PostgreSQL with pgvector for relational data and RAG with metadata filtering, and Azure Managed Redis for lowest-latency caching and vector indexing. Keep vector search firmly distinct from full-text search.

Also learn the optimization levers: Cosmos DB indexing and consistency, pgvector indexing and resource sizing, and connection pooling. Build a small RAG pipeline in a free subscription to make it concrete. Afterwards, move to Domain 3, connecting to and consuming Azure services. Track missed practice questions by data service, so you can see whether Cosmos DB, PostgreSQL or Redis is the weak spot.

FAQ

How much of AI-200 is Domain 2?

Develop AI solutions by using Azure data management services is 25–30% of the exam, the single largest domain and the biggest difference from AZ-204.

What is vector similarity search?

Finding the items whose embedding vectors are closest to a query's embedding, which retrieves semantically similar content rather than keyword matches.

What is retrieval-augmented generation (RAG)?

A pattern that retrieves relevant context, often by vector similarity search with metadata filtering, and grounds a language model's response on that retrieved context.

How do I run vector search in PostgreSQL?

Install and use the pgvector extension to store embeddings and run vector similarity search, with indexing and metadata filtering for performance and relevance.

What controls cost in Cosmos DB?

Request Units. Indexing policies, query design and the chosen consistency level all affect RU consumption, so they are the main cost and performance levers.

Can Azure Managed Redis do vector search?

Yes. Beyond caching with expiration and invalidation, Redis supports vector indexing for similarity search, serving as a low-latency vector store.

How does Domain 2 connect to the rest of the exam?

These AI data services are consumed by the containerized services of Domain 1, triggered through the messaging of Domain 3, and secured and monitored in Domain 4.

Other Cloud Computing certifications

Where the same subject shows up in other exams. Cards marked "Same subject" open the matching domain.

Test your Domain 2 knowledge

You have covered vector search and RAG across Cosmos DB, PostgreSQL and Redis. Test whether you can choose the right data service and retrieval approach under exam conditions.

Take the AI-200 practice test Check your AI-200 readiness

Keep going