EnglishDeutschFrançaisEspañolPortuguês

Databricks · DB-GAIE · Associate

Databricks Generative AI Engineer Associate — Practice Questions and Mock Exam

Prepare for DB-GAIE with original practice questions and clear answer explanations. Ask Alex, your AI tutor, when you need more detail, use your results to identify topics to review, and practice your pacing with timed mock exams.

Start practicing DB-GAIE freeBuy DB-GAIE access for $29.99
Buy DB-GAIE access for $29.99Start practicing DB-GAIE free
45Mock exam questions
90minTime limit

Checked against Databricks · August 2026 · Current exam version

Overview

About the exam

The Databricks Certified Generative AI Engineer Associate exam validates the ability to design, build, and deploy generative AI applications on the Databricks platform. It covers application design, data preparation and chunking, assembling and deploying retrieval-augmented generation pipelines, evaluation and monitoring, and governance. Candidates must demonstrate proficiency in prompt engineering, model-serving endpoints, and Databricks' generative AI tooling.

This certification is designed for AI engineers and developers with six or more months of experience building generative AI applications on the Databricks platform. It demonstrates the ability to take a generative AI application from design through production deployment.

Try the practice preview

Try five DB-GAIE questions

Try five practice questions from the app’s current Databricks Certified Generative AI Engineer Associate question bank, with answers and explanations.

Governance1 / 5

What is the relationship between a Databricks workspace and Unity Catalog metastore in the context of GenAI governance?

AlexFull explanation from Alex

In Unity Catalog, a metastore is the top-level container holding the catalog → schema → asset hierarchy with permissions. Multiple workspaces in the same region can share a single metastore, enabling cross-workspace governance of models, tables, and other assets. This means a model registered in dev is immediately visible in production with consistent access controls—no copying required. Distractor analysis: The option “A metastore can attach to at most one workspace at a time…” is wrong because metastores attach to many workspaces, not one. The option “Each workspace requires a dedicated metastore…” is wrong because shared metastores eliminate the need for per-environment metastores. The option “Unity Catalog assigns one metastore per cloud region…” is wrong because Unity Catalog assigns one metastore per region but workspaces within that region do share governance. Ref: docs.databricks.com/en/data-governance/unity-catalog/metastores.html

Sourcedocs.databricks.com

Assembling and Deploying Applications2 / 5

A team wants to A/B test two different prompt strategies for their customer support agent. How should they set this up using Databricks serving?

AlexFull explanation from Alex

A/B testing GenAI agents on Databricks uses the traffic_config in a serving endpoint to split requests between served entities. Each agent version is deployed as a separate served_entity on the same endpoint, with traffic_percentage controlling the split (must sum to 100). Each version's responses are logged separately in the inference table, enabling comparison of quality, latency, and user feedback. Distractor analysis: The option “Use Feature Serving to tag each request and route…” is wrong—Feature Serving is for feature lookup, not prompt routing. The option “Deploy a single endpoint with a pre-processing hook…” is wrong—pre-processing hooks don't provide proper traffic management or independent logging. The option “Register each prompt version in MLflow Model Registry…” is wrong—custom routing notebooks bypass Databricks' built-in traffic management and inference logging. Ref: docs.databricks.com/en/machine-learning/model-serving/serve-multiple-models-to-serving-endpoint.html

Sourcedocs.databricks.com

Design Applications3 / 5

Which approach allows a GenAI agent to access the most recent data when answering questions about current events?

AlexFull explanation from Alex

LLM knowledge is frozen at training time, so it cannot answer questions about current events. A RAG pipeline with Delta Live Tables (the option “A retrieval agent querying a continuously refreshed Delta…”) only retrieves indexed documents—not live web data. Expanding the context window with cached snapshots (the option “Expanding the context window to include a cached snapshot of…”) still uses stale data. Weekly fine-tuning (the option “Fine-tuning the base model weekly…”) introduces a multi-day lag. A tool-calling agent with web search or real-time API tools (the option “A tool-calling agent with a web search or real-time API tool”) invokes external data sources on demand, returning up-to-the-minute results. Databricks supports this via MCP servers, custom tools, and built-in web search integration with Gemini/OpenAI/Anthropic models. Ref: docs.databricks.com/en/machine-learning/model-serving/web-search; docs.databricks.com/en/generative-ai/agent-framework/agent-tool.

Sourcedocs.databricks.com

Data Preparation4 / 5

When creating a Delta Sync vector search index, what parameter specifies the column containing text to be automatically embedded?

AlexFull explanation from Alex

When creating a Delta Sync vector search index with managed embeddings, the embedding_source_column parameter identifies the text column Databricks will automatically embed. The platform calls the specified embedding model endpoint, batches the text, and stores resulting vectors in the index. The options “auto_embedding_column”, “vector_embedding_source”, and “text_vectorization_col” are fabricated parameter names—none appear in the Databricks SDK or REST API. For self-managed embeddings, you instead specify an embedding_vector_column containing pre-computed vectors. Ref: docs.databricks.com/en/generative-ai/vector-search/create-index; docs.databricks.com/en/vector-search/query-vector-search.

Sourceapi-docs.databricks.com

Evaluation and Monitoring5 / 5

Which Databricks feature allows engineers to trace the complete execution flow of a GenAI agent, including tool calls, retrieval steps, and LLM interactions?

AlexFull explanation from Alex

MLflow Tracing instruments GenAI applications automatically when using supported frameworks (LangChain, OpenAI, Databricks SDK). Each request creates a trace containing hierarchical spans that represent individual operations. Engineers can view traces in the MLflow UI to debug issues, identify bottlenecks (which tool call is slow?), verify retrieval quality (what did the retriever return?), and understand agent reasoning (what did the LLM decide and why?).

Sourcedocs.databricks.com

Question pool

326 practice questions

The Pass-IT question pool gives you material to practice for DB-GAIE. A Pass-IT mock exam uses 45 questions and a 90-minute time limit; these are practice settings.

Pool details: DB-GAIE

Objectives in the guide56 objectives listed in the official guide

across 6 domains in the official exam guide

Pool size326 questions

= The pool size is equivalent to 7 sets of 45 questions; this does not mean that each mock exam uses a separate set.

Blueprint domains6 domains in the exam blueprint

Application Development 85 · Assembling and Deploying Applications 59 · Design Applications 47 · Data Preparation 51 · Evaluation and Monitoring 47 · Governance 37

Recorded as checked against sources326 of 326

questions recorded as having their answer, options, and explanation checked against official Databricks documentation

How questions are made →

Exam domains

What's on the exam

Application development is the dominant domain at 30%, nearly a third of the exam, covering framework choices like LangChain, prompt engineering, LLM guardrails, and model selection from a hub or marketplace. Assembling and deploying applications follows at 22% — coding chains as pyfunc models, registering them to Unity Catalog, and standing up Vector Search indexes for retrieval.

Design applications and data preparation each carry 14%, testing prompt design against business requirements and chunking strategy against document structure. Evaluation and monitoring (12%) and governance (8%) round out the exam, covering MLflow scoring, AI Gateway tracking, and guardrails against malicious input or licensing risk in source data.

Exam blueprint: DB-GAIE

Application Development30%

Assemble a GenAI application from framework choices like LangChain, a chunking and embedding strategy tuned to the source documents, and prompts that get augmented with user context, then guard against bad outputs, pick models using metrics and model cards, and track the whole system with MLflow and the Agent Framework.

≈ 24 h
Assembling and Deploying Applications22%

Code and deploy chains using pyfunc models with pre- and post-processing, assemble RAG applications with the appropriate embedding model, retriever, and Vector Search index, and register models to Unity Catalog with MLflow. Also covers applying CI/CD best practices, integrating MCP servers, and building user-facing interfaces for agent applications.

≈ 18 h
Design Applications14%

Craft prompts that produce a specific output shape, pick the model tasks and chain components a business requirement calls for, translate that requirement into concrete pipeline inputs and outputs, and decide when Agent Bricks like Knowledge Assistant or Multiagent Supervisor are the right building block.

≈ 11 h
Data Preparation14%

Chunk and clean source documents so only the useful content feeds a RAG pipeline, land that text in Delta tables under Unity Catalog, and judge retrieval quality well enough to know when re-ranking would help.

≈ 11 h
Evaluation and Monitoring12%

Choose an LLM and the metrics worth tracking for a given deployment, score and trace agent or RAG performance in MLflow, watch a live endpoint through AI Gateway and Agent Monitoring, and fold subject-matter feedback back into the system to keep improving it.

≈ 10 h
Governance8%

Mask sensitive fields to hit performance targets, guard the application against malicious user input, and check data sources against legal and licensing risk before problematic content ever reaches the model.

≈ 6 h

Format

Exam format and question types

The exam consists of 45 scored questions, a mix of multiple-choice and multiple-select formats, inside a 90-minute window; unscored pilot items may also appear. Questions run scenario-heavy: picking a chunking approach that fits a document set, matching chain components to a target input and output shape, or configuring Vector Search for a given latency and cost budget.

Question types: DB-GAIE

Multiple Choice100%

Select the single answer that best meets the question’s requirements.

See Databricks for official question-format information. The shares shown describe the Pass-IT practice pool; they do not establish the proportions on the official exam.

Planning

Preparing for DB-GAIE

The exam is delivered online with a remote proctor or at a test center, offered in English, Japanese, Portuguese (BR), and Korean. The credential holds for two years, and recertifying means passing the current version of the exam.

Preparation and logistics: DB-GAIE

Preparation

Illustrative study time50–120 h

illustrative planning range: 50 h with relevant experience to 120 h when starting out; your needs may fall outside this range

LevelAssociate
Recommended backgroundNone required. 6+ months hands-on GenAI development on Databricks recommended.

Taking and maintaining the certification

DeliveryOnline proctored or test center
Retake policyNo mandatory waiting period. Retake fee applies.
Certification validity2 years

Recertification required every 2 years by taking the current exam version.

Official guideView on provider site

Study notes

Common pitfalls

Topics to review: DB-GAIE

  1. 01Agent Bricks Selection

    Building a custom multi-agent chain from scratch instead of recognizing when a prebuilt Agent Bricks template for grounded Q&A, multi-agent routing, or structured extraction already solves the scenario leads to wrong answers on application-design questions.

  2. 02MCP Server Integration

    Not distinguishing a managed MCP connector from an externally hosted or fully custom MCP server for a given application requirement leads to wrong answers on tool-and-agent assembly questions.

  3. 03Re-ranking Role

    Skipping re-ranking as a distinct retrieval step, instead of treating the top vector-search results as final, leads to wrong answers on retrieval-pipeline design questions.

  4. 04AI Gateway Tracking

    Confusing AI Gateway's inference and usage tracking with basic inference-table logging, or forgetting rate limiting as a cost control, leads to wrong answers on LLM-deployment monitoring questions.

  5. 05Licensing Risk

    Overlooking data-source licensing terms when selecting training or retrieval content leads to wrong answers on governance questions about legal risk in a GenAI pipeline.

Practice with questions and explanations in Pass-IT →

Questions

Frequently asked questions

How long is the Databricks Certified Generative AI Engineer Associate exam?

The Databricks Certified Generative AI Engineer Associate exam has 45 questions and a 90-minute time limit.

Which pitfalls should I review when preparing for Databricks Certified Generative AI Engineer Associate?

Topics to review include Agent Bricks Selection, MCP Server Integration, Re-ranking Role, AI Gateway Tracking, Licensing Risk. Work through examples to check that you understand the distinctions and can explain your answer.

Which Generative AI Engineer Associate topics carry the most weight?

Application development is the largest area at 30%, followed by assembling and deploying applications at 22%. Design and data preparation take 14% each, evaluation and monitoring 12% and governance 8%. More than half the exam is therefore building and shipping a generative application rather than reasoning about models.

What do you need before the Generative AI Engineer Associate exam?

No certification is required. Databricks recommends six or more months of hands-on generative AI development on the platform, and the catalog budget is about 80 hours. The exam is written from the builder side, so retrieval, chaining and deployment questions assume you have shipped something.

How long does the Generative AI Engineer Associate last?

Two years, after which you sit the current version of the exam again. Databricks offers no continuing-education alternative. On a subject moving this fast, the resit is likely to be a materially different exam.

Generative AI Engineer or Machine Learning Associate?

The machine learning exam is about training, tracking and deploying your own models, with Databricks Machine Learning itself at 38% of it. The generative AI exam assumes the model already exists and tests what you build around it: retrieval, chaining, evaluation and governance. Take the one that matches whether you train models or wire them up.

One certification, 12 months

Practice for DB-GAIE

Focus your practice on one certification, or choose Pro to practice across all certifications.

Buy DB-GAIE access for $29.99
Start a free practice sessionTry the first 20 questions without a card to see whether the practice suits you.

For eligible purchases: money-back guarantee if you fail your exam.

View guarantee terms →

Related certifications

Databricks Certified Associate Developer for Apache SparkDatabricks Certified Data Analyst AssociateDatabricks Certified Data Engineer AssociateDatabricks Certified Data Engineer ProfessionalAll Databricks certifications →