Databricks GenAI Engineer Associate
Databricks certificationGenerative AI Engineer Associate

Pass the Databricks GenAI Engineer Associate

Study guide, teaching cards, practice banks and a full mock for the Databricks Certified Generative AI Engineer Associate. Built to the official exam guide, with bridges from what you already know from Azure AI-103 and Claude CCA-F.

The exam (verified 1 Nov 2026)
ExamDatabricks Certified Generative AI Engineer Associate (proctored)
Exam guideTitled "Exam Guide – August 2026" and linked from the certification page (file path dated 2026-10). Its updates table says: exam published 18 March 2026, guide updated 18 March 2026.
Scored items45 scored items, multiple choice and multiple selection
Unscored itemsUp to 10 more unscored items, not identified, which don't count
Time90 minutes, with extra time for unscored items already included
Passing scoreNot published in the guide or on the certification page. This site's mock and banks use a self-set target of 80%.
Question typesMultiple choice (choose one) and multiple selection ("Which TWO…"). No case studies, drag and drop or code-completion items.
Test aidsNone allowed (closed book)
FeeUSD 200 plus tax
DeliveryGuide: online proctored. Certification page: online or test center.
LanguagesEnglish, Japanese, Brazilian Portuguese, Korean
PrerequisitesNone. Training is highly recommended, along with 6+ months of hands-on experience with the outline's tasks.
Validity2 years. You recertify by taking the current live exam.
Recommended trainingInstructor-led: Generative AI Engineering with Databricks. Self-paced (Databricks Academy): Building RAG Agents with Agent Bricks; Building Agentic Applications on Databricks; Agent Evaluation on Databricks; Deploying and Monitoring Agent Applications on Databricks.

Sources: certification page · official exam guide (PDF). The guide asks you to check it again two weeks before your exam, because the outline can change.

Docs have moved past the exam guide. The guide is from March 2026, but the Databricks docs changed several names in the months after it. The exam tests the guide's names, so this site uses them and notes the new names on the cards.

Sections and weights

SectionName (official)Weight≈ items of 45Here
Section 1Design Applications14%68 cards · 18 items
Section 2Data Preparation14%68 cards · 17 items
Section 3Application Development30%1410 cards · 20 items
Section 4Assembling and Deploying Applications22%1010 cards · 20 items
Section 5Governance8%45 cards · 16 items
Section 6Evaluation and Monitoring12%59 cards · 19 items

Weights from the official exam guide and certification page. The item counts are this site's rounding of 45 × weight; Databricks doesn't publish per-section counts.

Start here

S1 · 14%Design Applications

8 teaching cards and a 18-item practice bank.

S2 · 14%Data Preparation

8 teaching cards and a 17-item practice bank.

S3 · 30%Application Development

10 teaching cards and a 20-item practice bank.

S4 · 22%Assembling and Deploying Applications

10 teaching cards and a 20-item practice bank.

S5 · 8%Governance

5 teaching cards and a 16-item practice bank.

S6 · 12%Evaluation and Monitoring

9 teaching cards and a 19-item practice bank.

Mock · 90 minMock Exam #1

45 items weighted by section, exam-level stems, every option explained.

Services and features map

Only what the March 2026 outline covers. "Depth" is this site's reading of the objectives: High means expect several questions on details and trade-offs; Light means know what it's for.

Databricks pieceWhat it's forExam depthSections
Agent Bricks: Knowledge AssistantNo-code document Q&A agent with citations over UC volumes, tables or indexes; SMEs improve it with labelled questions and guidelines.Medium: when to use it vs build your ownS1
Agent Bricks: Supervisor Agent (Multi-Agent Supervisor)Coordinates subagents and tools, including Genie, agent endpoints and UC functions. Each end user needs access to each subagent.MediumS1
Agent Bricks: Information ExtractionTurns documents into JSON for a schema you define. Built on ai_extract; runs from SQL, a pipeline or a REST endpoint.MediumS1
Mosaic AI Vector Search (now AI Search)Managed vector index. Endpoint types: standard (about 320M vectors at 768 dims, high QPS) or storage-optimized (over 1B vectors, about 250 ms slower, triggered sync only). Index types: Delta Sync (managed or self-managed embeddings) or Direct Vector Access. Supports hybrid search, filters and a reranker.High: create, query, choose an endpoint, keep in syncS2, S4
Delta Lake + Unity CatalogChunk tables (Change Data Feed on for Delta Sync), volumes for raw files, three-level names, grants, models and functions in UC, row filters and column masks.HighS2, S4, S5
AI Functionsai_query (general batch inference in SQL), ai_parse_document (PDF, Office and images to elements), ai_extract, ai_mask, ai_classify, ai_summarize.Medium: batch inference with ai_query is an objectiveS2, S4, S5
Foundation Model APIsDatabricks-hosted LLMs and embedding models. Pay-per-token (including a priority tier) for prototyping, provisioned throughput for production, fine-tuned weights and HIPAA.HighS3, S4
Model ServingEndpoints for agents, custom models and external models. Permissions are CAN VIEW, CAN QUERY and CAN MANAGE. Resource access works through automatic passthrough (declared resources), on-behalf-of-user (preview) or secrets.HighS4
MLflow models: pyfunc and models from codeWrap a chain with pre- and post-processing, set a signature and input example, log dependencies, and register to UC with aliases (no stages).HighS4
Mosaic AI Agent FrameworkAuthor agents with MLflow ResponsesAgent, tools from UC functions, retriever tools, and deploy with agents.deploy(). Deploying also sets up a Review App, tracing and inference tables.HighS3, S4
LangChain on Databricksdatabricks-langchain classes: ChatDatabricks, DatabricksEmbeddings, DatabricksVectorSearch retriever, UCFunctionToolkit, GenieAgent.MediumS3
Genie Spaces (now Genie Agents) and Conversation APINatural-language questions over tables. The API returns SQL and results and applies the caller's UC permissions. In multi-agent systems, Genie is the structured-data worker.MediumS3
MLflow Prompt Registry (Beta)Immutable prompt versions, movable aliases such as @production, loaded by URI. This is how you promote prompts across environments and roll back.MediumS4
MCP serversManaged (Databricks-hosted: Genie, Vector Search, UC functions, SQL), external (third-party, registered through a UC connection with secrets) and custom (your own, hosted on Databricks Apps).MediumS4
Agent memory (Lakebase)Sessions keep one conversation; memory keeps facts across conversations. Both are managed and backed by Lakebase.Light to mediumS4
Databricks Apps, Review App, Slack/TeamsUser-facing interfaces. Apps keep tokens server-side and use the signed-in user's identity. The Review App collects SME feedback. Genie integrates with Teams and Slack.MediumS4, S6
MLflow evaluation: judges and scorersmlflow.genai.evaluate() with built-in judges. Correctness, RetrievalSufficiency and ToolCallCorrectness need ground truth; RelevanceToQuery, RetrievalRelevance, RetrievalGroundedness, Safety and Guidelines don't. Custom @scorer functions; make_judge and judge alignment from SME labels.HighS6
MLflow TracingRecords every step's inputs, outputs, latency and tokens. Autolog covers common frameworks, and @mlflow.trace covers your own functions.MediumS6
Production monitoring (Agent Monitoring)The same scorers run on a sample of live traces (register() then start(sampling_config=...)). Beta.MediumS6
AI Gateway (legacy; now Unity Gateway)On serving endpoints: inference tables (requests and responses to Delta), usage tracking (system.serving.endpoint_usage), rate limits (QPM/TPM by endpoint, user or group), guardrails (safety filter; PII Block or Mask; preview).HighS3, S5, S6
Cost controlsPay-per-token vs provisioned throughput, scale to zero, rate limits, smaller models, batch with ai_query, usage system tables. Serverless budget policies only tag usage; they don't cap spend.MediumS6
Open-source parserspypdf / PyPDF2 (PDF text), unstructured (mixed formats), Tesseract / pytesseract (OCR on scans), BeautifulSoup (HTML).Light: pick the right package for the formatS2

Two-week study plan

About 90 minutes a day, starting the day after AI-103. S3 and S4 together are 52% of the exam, so they get four days. Where Azure or Claude already taught the concept, read the bridge line on the card and spend the time on the Databricks names.

DayStudyPractice
Day 1Re-read the official exam guide (it asks for this two weeks before the exam). Read this hub's map and drift note, then the S1 cards.S1 practice bank
Day 2S2 cards: chunking, parsers, writing chunks to Delta, retrieval metrics, reranking.S2 practice bank
Day 3S3 clusters A and B: frameworks, choosing models, prompts, guardrails. Skim anything you already know from AI-103 and CCA-F.First half of S3 bank
Day 4S3 cluster C: Agent Framework, evaluation vs monitoring, Genie in multi-agent systems.Rest of S3 bank
Day 5S4 clusters A and B: pyfunc, signatures, UC registration, Vector Search endpoints and indexes.S4 items on these cards
Day 6S4 cluster C: serving access, Foundation Model APIs, ai_query, MCP, memory, prompt registry, CI/CD, interfaces.Rest of S4 bank
Day 7Hands-on in Databricks Free Edition (the guide recommends it): parse a few PDFs, chunk to a Delta table, build a Vector Search index, query it, then batch-summarise with ai_query.Write down what surprised you
Day 8S5 and S6 cards: masking, guardrails, licensing; judges, scorers, tracing, inference tables, AI Gateway, cost.S5 bank
Day 9Learn the judges table by heart: which need ground truth, which don't.S6 bank
Day 10Mock Exam #1 in exam mode: 90 minutes, no notes.Score it; list every miss
Day 11For each mock miss, reread its card and write down the rule you missed.Redo the missed items in practice mode
Day 12Hands-on: log a small agent with ResponsesAgent, register it to UC with an alias, run mlflow.genai.evaluate() with two built-in judges and one custom scorer.Re-run S3 and S4 banks in exam mode
Day 13Weakest two sections again. Re-run their banks in exam mode until you score 80% or more.S5 and S6 banks in exam mode
Day 14Light day: drift note, judges table, Vector Search endpoint table. Retake Mock #1 in exam mode only if day 10 was under 80%.Rest
About the questions: every practice item is original, written to the exam guide's objectives. None is a real exam question, and none comes from a dump. Each has a deliberate runner-up and a plain-language explanation, and the build checks that "pick the longest option" or "pick the shortest option" score near chance and that answer positions are balanced.