Databricks Certified Generative AI Engineer Associate Study Guide

Domain 1: Design Applications

Key concepts you must know · 96 practice questions

Design a prompt that elicits a specifically formatted response, controlling structure, tone, and length.
Select the model task that fits a business requirement: summarization, classification, extraction, question answering, translation, or text generation.
Choose the chain components for a desired input and output: retriever, prompt template, LLM, output parser, and tools.
Translate a business goal into the concrete inputs and outputs of an AI pipeline.
Define and order the tools an agent uses to gather knowledge or take actions for multi-stage reasoning.
Decide when to use Agent Bricks: Knowledge Assistant for grounded question answering, Multiagent Supervisor to orchestrate multiple agents, and Information Extraction for structured fields from unstructured documents.

Domain 2: Data Preparation

Key concepts you must know · 95 practice questions

Apply a chunking strategy for a document structure and model constraints, balancing chunk size against overlap and the resulting embedding record count.
Filter extraneous content such as boilerplate, navigation, and markup that degrades RAG quality.
Pick the right Python package for the source format: OCR (pytesseract) for scanned images, BeautifulSoup for HTML, PDF text extractors for PDFs.
Write chunked text into Delta Lake tables in Unity Catalog as the basis for embedding and Vector Search sync.
Identify which source documents provide the knowledge and quality a RAG application needs.
Evaluate retrieval performance with metrics such as recall@k, precision, and context relevance.
Use advanced chunking (parent-document, sentence-window, metadata-enriched) and understand the role of re-ranking: retrieve broadly, then re-rank for precision.

Domain 3: Application Development

Key concepts you must know · 176 practice questions

Augment a prompt with context from the user's input based on key fields, terms, and intent.
Adjust an LLM response from a baseline to a desired output with few-shot examples, system prompts, and constraints.
Qualitatively assess responses for hallucination, unsafe content, and format drift.
Implement guardrails that filter input and output, restrict topics, and redact PII, including AI Gateway guardrails.
Select the best LLM by latency, cost, context window, and quality, and choose Foundation Model APIs versus External Models.
Select an embedding model and context length for the source documents and query patterns, trading dimension and quality against cost.
Build agentic systems with MLflow and the Mosaic AI Agent Framework, iterating in the AI Playground, and let agents use Genie spaces or the conversational API to retrieve governed data.

Domain 4: Assembling and Deploying Applications

Key concepts you must know · 152 practice questions

Code a chain as an MLflow pyfunc model with pre-processing and post-processing steps.
Identify the elements a RAG application needs: model flavor, embedding model, retriever, dependencies, input examples, and model signature.
Register the model or agent to Unity Catalog with MLflow using the three-level namespace catalog.schema.model.
Create and query a Mosaic AI Vector Search index, and configure it (standard versus storage-optimized, hybrid search, re-ranking) for the embedding count, update frequency, latency, and cost.
Serve applications with Foundation Model APIs (pay-per-token or provisioned throughput) and run batch inference with ai_query().
Integrate managed, external, and custom MCP servers, promote prompts across environments with MLflow versions and aliases, and control access to serving endpoints.
Build a user-facing interface with Databricks Apps, Slack, or Teams where the backend calls the agent endpoint with the app's identity and never exposes a token in the browser.

Domain 5: Governance

Key concepts you must know · 84 practice questions

Use masking as a guardrail to detect and redact PII before it reaches the model or the user.
Select guardrail techniques that protect against prompt injection and other malicious input.
Respect legal and licensing requirements for the data sources that ground an application.
Recommend alternatives that mitigate problematic (toxic, biased, or non-compliant) source text.
Lean on Unity Catalog permissions and lineage and AI Gateway guardrails to enforce governance on serving endpoints.

Domain 6: Evaluation and Monitoring

Key concepts you must know · 108 practice questions

Select an LLM by quantitative evaluation metrics and choose the key metrics to monitor for a deployment: latency, cost, quality, groundedness, and drift.
Evaluate agents with MLflow scoring and tracing using mlflow.genai.evaluate and custom scorers.
Use inference logging and inference tables with Agent and Lakehouse Monitoring to track a live endpoint.
Control LLM costs with provisioned throughput, caching, rate limiting, and right-sizing the model.
Distinguish evaluation judges that need ground truth from LLM-as-judge metrics, and use AI Gateway (inference tables, usage tables, rate limiting) to track deployed agents.
Incorporate subject-matter-expert feedback with clear rubrics and calibration rather than treating an LLM judge as ground truth.

Databricks Certified Generative AI Engineer Associate exam tips

Weight your prep toward Application Development (30%) and Assembling and Deploying (22%): together they are more than half the exam. Be fluent with the Mosaic AI Agent Framework, MLflow pyfunc chains, registering models to Unity Catalog, Vector Search index creation and configuration, Foundation Model APIs, and ai_query() for batch inference.
This is a Databricks-specific exam. When a generic answer and a Databricks answer both look right, the Databricks-native one usually wins: models in Unity Catalog (not the legacy workspace registry), Vector Search (not a bolt-on vector database), and prompt promotion via MLflow aliases (not overwriting tables or merging Git branches).
Know the batch-versus-real-time split cold: ai_query() is for batch scoring over a table, while a Model Serving endpoint handles low-latency single requests. Choosing the wrong one is a common trap.
Understand the three MCP server types (managed, external, custom) and the secure Databricks Apps pattern where the backend calls the agent with the app's identity and no token is exposed in browser JavaScript.
For evaluation, remember mlflow.genai.evaluate with built-in and custom scorers, the difference between judges that require ground truth and LLM-as-judge, and using inference tables plus AI Gateway to monitor cost and quality in production.

Study guide FAQ

How many questions are on the exam and is there a fixed passing score?

There are 45 scored questions (multiple choice and multiple response) in 90 minutes. Databricks does not publish a fixed passing percentage for this exam; you receive a pass or fail result. The certification is valid for two years and recertification requires taking the current live exam.

Do I need to write code to pass?

You should be comfortable reading and reasoning about Python, LangChain, and MLflow code, since questions are scenario-based around building RAG and agent applications. The exam tests your ability to choose the right approach and Databricks feature rather than to write large programs from scratch.

How is this different from the Machine Learning Associate exam?

The Generative AI Engineer Associate focuses on LLM and agent applications: prompts, RAG, Vector Search, the Agent Framework, guardrails, and LLM evaluation and monitoring. The Machine Learning Associate exam focuses on classic end-to-end ML with AutoML, scikit-learn, Spark ML, and model deployment.

Which Databricks tools should I know best?

The Mosaic AI stack: Agent Framework and Agent Evaluation, Vector Search, Model Serving, Foundation Model APIs, AI Gateway, the AI Playground, and Agent Bricks, together with Unity Catalog governance and MLflow (models in Unity Catalog, pyfunc, tracing, custom scorers, and prompt aliases).

Is CertGrid practice official Databricks material?

No. CertGrid is an independent practice platform and is not affiliated with or endorsed by Databricks. These questions are original and written to mirror the current exam guide so you can rehearse the objectives. Always confirm the current exam guide on the Databricks certification site before your exam.

Related Data resources

Test yourself: 711 Databricks Certified Generative AI Engineer Associate practice questions