DBX-GAI Databricks Generative AI Engineer Associate
- Vendor
- Databricks
- Level
- Associate
- Questions
- 45
- Time
- 90 min
- Domains
- 6
- In the bank
- 285
Practice 285 questionsTake a timed mock examDay-by-day study plan
Guides for this exam
- Prompting, RAG, Fine-Tuning or Continued Pre-Training: One Rule That Works on AWS, Azure, Google Cloud and Databricks Exams · 한국어
- DBX-GAI Study Guide: Read Every Question as a Design Review With Five Questions · 한국어
- 8 Mistakes That Repeat Across Every Databricks Exam, and How to Spot Each One in the Question · 한국어
The exam at a glance
| Level | Associate |
|---|---|
| Questions | 45 |
| Time | 90 minutes (about 2 minutes per question) |
| Passing mark | Not published by the vendor |
Domains and weights
| Domain | Weight | Practice questions here |
|---|---|---|
| D1 · Design Applications | 14% | 40 |
| D2 · Data Preparation | 14% | 39 |
| D3 · Application Development | 30% | 86 |
| D4 · Assembling and Deploying Applications | 22% | 63 |
| D5 · Governance | 8% | 23 |
| D6 · Evaluation and Monitoring | 12% | 34 |
47 of the 285 practice questions (16%) ask for more than one answer.
D1 · Design Applications 14%
Design questions ask which shape of generative AI application fits a requirement: plain prompt, RAG, agent with tools or multi-agent. The deciding detail is how fresh the knowledge must be and who decides the next step. Strong options use the simplest pattern that meets the need.
What to watch for
- Knowledge that changes often belongs in retrieval with source citations, not in fine-tuning.
- A LangChain output parser turns a chat message into a plain string or structured output.
- When the fields to extract change per request, pass them in the prompt instead of fixing a schema.
Tasks in the official exam guide (6)
- 1.1 Design a prompt that elicits a specifically formatted response
- 1.2 Select model tasks to accomplish a business requirement
- 1.3 Select chain components for a desired model input and output
- 1.4 Translate business goals into desired inputs and outputs of the AI pipeline
- 1.5 Define and order tools that gather knowledge or take actions for multi-stage reasoning
- 1.6 Decide how and when to use Agent Bricks (Knowledge Assistant, Multi-Agent Supervisor, Information Extraction)
D2 · Data Preparation 14%
Data preparation questions are about chunking, parsing and judging retrieval for RAG. Look at the document structure and the metric asked for. Strong options keep meaning together in a chunk and compute precision and recall from the right counts; weaker ones split by characters regardless of structure.
What to watch for
- Precision is relevant retrieved over retrieved; recall is relevant retrieved over all relevant in the index.
- Tables and sections should be parsed as units, not cut by a fixed character count.
- Overlap or small-to-big retrieval keeps the neighbouring sentences that explain an exception.
Tasks in the official exam guide (8)
- 2.1 Apply a chunking strategy for a document structure and model constraints
- 2.2 Filter extraneous content in source documents that degrades RAG quality
- 2.3 Choose the Python package to extract content from given source data and format
- 2.4 Define the operations and sequence to write chunked text into Delta tables in Unity Catalog
- 2.5 Identify source documents that provide the needed knowledge and quality for a RAG application
- 2.6 Use tools and metrics to evaluate retrieval performance
- 2.7 Design retrieval systems using advanced chunking strategies
- 2.8 Explain the role of re-ranking in information retrieval
D3 · Application Development 30%
Application development questions trace a RAG chain or agent and ask what to fix. Use the trace: check whether retrieval found the right passages before blaming the model. Strong options fix the stage that failed; weaker ones change the prompt when retrieval brought the wrong text.
What to watch for
- Wrong passages for a question are a retrieval problem: chunking, embeddings or filters, not the LLM.
- Model cards state languages, context length and license, which decide fit before any test.
- Genie Chat mode APIs are asynchronous: poll the message status until the query result is ready.
Tasks in the official exam guide (13)
- 3.1 Select LangChain or similar tools for a generative AI application
- 3.2 Qualitatively assess responses for common quality and safety issues
- 3.3 Select a chunking strategy based on model and retrieval evaluation
- 3.4 Augment a prompt with context from the user's input (key fields, terms, intents)
- 3.5 Create a prompt that moves an LLM's response from a baseline to the desired output
- 3.6 Implement LLM guardrails to prevent negative outcomes
- 3.7 Select the best LLM based on application attributes
- 3.8 Select an embedding model context length from source documents, expected queries and optimization strategy
- 3.9 Select a model from a model hub or marketplace using model metadata and model cards
- 3.10 Select the best model for a task from common experiment metrics
- 3.11 Use MLflow and Agent Framework to develop agentic systems
- 3.12 Compare the evaluation and monitoring phases of the Gen AI application lifecycle
- 3.13 Enable multi-agent systems to use Genie spaces or the Conversation API to retrieve data
D4 · Assembling and Deploying Applications 22%
Assembling and deploying questions cover logging agents with MLflow, packaging files, authentication and vector search. The deciding detail is what the serving environment can reach. Strong options log everything the model needs as artifacts or resources; weaker ones rely on paths that exist only in development.
What to watch for
- Log files a chain reads as model artifacts so the endpoint has them; development volume paths may not be reachable.
- Declare resources such as indexes and UC functions when logging, so serving gets automatic short-lived credentials.
- A Direct Vector Access index takes writes through the API without a source Delta table.
Tasks in the official exam guide (15)
- 4.1 Code a chain using a pyfunc model with pre- and post-processing
- 4.2 Control access to resources from model serving endpoints
- 4.3 Code a simple chain according to requirements
- 4.4 Choose the basic elements of a RAG application: model flavor, embedding model, retriever, dependencies, input examples, signature
- 4.5 Register the model to Unity Catalog using MLflow
- 4.6 Create and query a Vector Search index
- 4.7 Identify how to serve an LLM application that uses Foundation Model APIs
- 4.8 Explain the key concepts and components of Vector Search
- 4.9 Identify batch inference workloads and apply ai_query() appropriately
- 4.10 Configure Vector Search for embedding count, update frequency, latency and cost requirements
- 4.11 Configure a persistent datastore for intermediate memory or structured information
- 4.12 Apply CI/CD practices: update a Vector Search index, promote prompts across environments, test agent components
- 4.13 Integrate managed, external and custom MCP servers for application requirements
- 4.14 Apply prompt version control and manage the prompt lifecycle
- 4.15 Develop an interactive user-facing interface for an agent (Apps, Slack, Teams, etc.)
D5 · Governance 8%
Governance questions ask how to keep an assistant safe, compliant and affordable. Read the constraint: latency budget, license end date, data type to mask. Strong options combine fast pattern rules with targeted controls and remove data at the source; weaker ones add heavy checks or leave copies behind.
What to watch for
- Fixed patterns like card numbers mask fast with rules; names need an entity model.
- Removing licensed content means deleting the rows and syncing the index, not just filtering answers.
- Rate limits and input size limits guard an endpoint against cost and abuse.
Tasks in the official exam guide (4)
- 5.1 Use masking techniques as guardrails to meet a performance objective
- 5.2 Select guardrail techniques that protect against malicious user inputs
- 5.3 Use legal and licensing requirements of data sources to avoid legal risk
- 5.4 Recommend an alternative for problematic text in a data source feeding a GenAI application
D6 · Evaluation and Monitoring 12%
Evaluation and monitoring questions ask how to measure an application and watch it in production. Prefer evidence from your own task and traces over public rankings. Strong options use the tracing, judges and limits that match the need; weaker ones choose models or budgets by general benchmarks.
What to watch for
- Rank candidate models on an evaluation set from your own use case, not a public leaderboard.
- MLflow traces from deployed agents let developers step through individual production requests.
- Reviewers in the Review App need workspace access and permission on the experiment holding the traces.
Tasks in the official exam guide (10)
- 6.1 Select an LLM (size and architecture) from quantitative evaluation metrics
- 6.2 Select key metrics to monitor for an LLM deployment scenario
- 6.3 Evaluate agent performance using MLflow scoring and tracing
- 6.4 Use inference logging to assess a deployed RAG application
- 6.5 Use Databricks features to control LLM costs
- 6.6 Use inference tables and Agent Monitoring to track a live LLM endpoint
- 6.7 Identify evaluation judges that require ground truth
- 6.8 Use AI Gateway (inference tables, usage tables, rate limiting) to track an LLM or agent deployed with Agent Framework
- 6.9 Use Databricks custom scorers to evaluate agents and LLMs
- 6.10 Incorporate SME feedback to improve agent performance
A study order that works
The 19-day study plan for DBX-GAI takes the domains in this order. Each day has a short reading, key terms and a quiz from the bank.
- Days 1–2: D1 · Design Applications
- Days 3–5: D2 · Data Preparation
- Days 6–9: D3 · Application Development
- Days 10–14: D4 · Assembling and Deploying Applications
- Day 15: D5 · Governance
- Days 16–18: D6 · Evaluation and Monitoring
- Day 19: Final Review: Decision Cheat-Sheet Across the Six Domains
Recent changes to these questions
- : 3 answer choice sets changed, 3 answer keys corrected, 3 explanations rewritten
- : 60 questions added
- : 1 answer choice set changed, 1 answer key corrected, 1 explanation rewritten
- : Exam added to the bank (225 questions added)
About these questions
Every question is original, written from the public exam guide and checked against the vendor's own documentation. None comes from a real exam. Each one has an explanation of why the answer is right and why the other options are wrong. How the questions are written.