Free study guide · Databricks
How to pass the Databricks Certified Generative AI Engineer Associate
What the exam covers, how to plan your study, and how to read its questions. From the CertLumen course for this exam, kept up to date with the official exam guide.
- Questions
- 45
- Time
- 90 minutes
- Passing score
- Not published by Databricks; this course targets ≥ 80% on unseen mocks
- Exam fee
- USD 200 (plus applicable taxes)
- Valid for
- 2 years (recertify by taking the current live exam)
- Delivery
- Online proctored
What the exam covers
A study plan that works
The default path for engineers who have built something with an LLM but not the full Databricks stack. It covers all 19 lessons and all 7 labs in section order, with a section drill after each section and daily flashcards. Application Development (g3, 30%) and Assembling and Deploying (g4, 22%) get the most days because together they carry more than half of the scored questions. Mock Exam 1 is on day 13 and Mock Exam 2 on day 14, so you sit the exam on day 15. Every lab runs on Databricks Free Edition except the steps marked as needing a paid workspace. Plan on about 1.5–2.5 hours a day.
| Day | Focus | Time |
|---|---|---|
| 1 | Kick-off: the exam, the platform changes and the I/O contract | 90 min |
| 2 | Chains, tools and Agent Bricks, then the g1 drill | 125 min |
| 3 | Source documents, extraction packages and chunking | 105 min |
| 4 | Advanced retrieval, re-ranking and the g2 lab | 135 min |
| 5 | Frameworks, prompt augmentation and guardrails | 135 min |
| 6 | Choosing the LLM and the embedding window, with the g3 lab | 140 min |
| 7 | Agentic systems and Genie, then the g3 drill | 155 min |
| 8 | pyfunc chains, Unity Catalog registration and Vector Search | 115 min |
| 9 | Index lab, serving and access control | 130 min |
| 10 | Memory, MCP, prompts and CI/CD, then the g4 drill | 145 min |
| 11 | Governance, then model choice, monitoring metrics and cost | 115 min |
| 12 | Evaluation, tracing and live monitoring | 165 min |
| 13 | g6 drill and Mock Exam 1 | 160 min |
| 14 | Mock Exam 2 and light review | 140 min |
Exam strategy
How the exam works
The Databricks Certified Generative AI Engineer Associate exam checks that you can design and build LLM applications on Databricks: break a requirement into tasks, prepare the documents, pick models and tools, assemble a RAG chain or an agent, deploy and govern it, then evaluate and monitor it. Most questions are short scenarios that end with "which approach" or "what should the engineer do". Some ask you to read a few lines of Python or SQL. Every wrong option is usually a real Databricks feature used in the wrong place.
| Fact | Detail (exam guide, exam page and certification FAQ) |
|---|---|
| Scored questions | 45, multiple choice and multiple selection |
| Unscored items | Up to 10 more, not marked and not scored. The time limit already includes extra time for them |
| Time limit | 90 minutes |
| Fee | USD 200 plus applicable taxes. Each retake costs the full fee |
| Delivery | Online proctored through Kryterion (Webassessor). The exam page also lists test centers |
| Languages | English, Japanese, Brazilian Portuguese, Korean |
| Test aids | None |
| Prerequisites | None. Related training and six months of hands-on experience are highly recommended |
| Passing score | Not published, because it can change. This course targets ≥ 80% on unseen mocks |
| Validity | Two years. You recertify by taking the current live exam |
| Exam version | The current guide (the PDF file is dated August 2026) lists the exam as published and updated on March 18, 2026. Check the guide two weeks before your date |
The exam mixes single-answer and multiple-selection questions, and so does this course: about one question in seven says "(Choose two.)". Judge each option on its own as true or false for the scenario, then pick exactly the number asked for. The official sample questions 1 and 9 are both choose-two.
Where the points are
The guide gives percentage weights for six sections. On 45 scored items they come to the counts below, and each of this course's two mock exams uses exactly those counts.
| Section | Weight | Questions per mock (of 45) | Highest-yield topics |
|---|---|---|---|
| g1 Design Applications | 14% | 6 | Input/output contracts, task selection, format prompts and structured outputs, tool ordering, Agent Bricks decision rule |
| g2 Data Preparation | 14% | 6 | Extraction packages by format, chunk size and overlap, writing chunks to a CDF-enabled Delta table, retrieval metrics, re-ranking |
| g3 Application Development | 30% | 14 | Framework choice, prompt augmentation, guardrails, model and embedding selection, evaluation vs monitoring, ResponsesAgent, Genie workers |
| g4 Assembling and Deploying | 22% | 10 | pyfunc chains, Unity Catalog registration, Vector Search endpoints and sync, authentication from serving endpoints, ai_query, MCP, Prompt Registry, CI/CD |
| g5 Governance | 8% | 4 | Mask vs Block, prompt-injection layers, data licenses, fixing problematic text upstream |
| g6 Evaluation and Monitoring | 12% | 5 | Choosing a model from metrics, judges that need ground truth, custom scorers, inference and usage tables, rate limits, SME feedback |
Application Development and Assembling and Deploying together are 52% of the score, about 24 questions. A weak g5 costs you at most four questions; a weak g3 can sink the exam. Spend your extra time where the weight is.
The exam guide versus the October 2026 platform
The exam guide was published on March 18, 2026. Since then Databricks has reorganised its agent stack, so the documentation you read today describes several exam topics under new names or labels them legacy. The exam still tests the workflow in the guide. Every lesson in this course handles the change the same way: it teaches the exam-era workflow first, then a short "October 2026" note names what replaced it. This table puts the whole picture in one place.
| Exam guide says | What the exam tests | October 2026 docs | Status |
|---|---|---|---|
| Mosaic AI Vector Search | Endpoints, Delta Sync and Direct Vector Access indexes, triggered and continuous sync, standard and storage-optimized endpoints | Databricks AI Search (formerly Vector Search). Python package databricks-ai-search with AISearchClient; UI under Compute > AI Search | Renamed. Same concepts and method names |
| Genie Spaces | Text-to-SQL worker for a multi-agent system, the Conversation API | Genie Agents (formerly Genie Spaces). The API keeps the old name in paths and fields (/genie/spaces, space_id, DatabricksGenieSpace) | Renamed |
| Agent Framework | Log with models from code and resources, register in Unity Catalog, deploy with agents.deploy() to Model Serving (Review App, tracing, inference tables, monitoring) | New agents: Agent Bricks CLI (agentbricks init, dev, deploy) and DurableAgentServer on Agent Runtime, which runs on Databricks Apps. Model Serving agents are labelled legacy | CLI in Beta. The Model Serving path is legacy but still works |
| Agent Bricks: Knowledge Assistant, Multiagent Supervisor | Which builder fits a scenario | Agent Bricks is now the name of the whole agent developer platform. Knowledge Assistant and Supervisor Agent are legacy; new low-code agents use Genie Agents and coded multi-agent systems use the Agent Bricks CLI | Legacy, not recommended for new agents |
| Agent Bricks: Information Extraction | Documents to structured rows | Rebuilt on ai_extract, run in SQL or a Lakeflow pipeline. The older Beta version is legacy | Current |
| AI Gateway (inference tables, usage tables, rate limiting) | Which gateway features each endpoint type supports. Agent endpoints get payload logging only | Labelled AI Gateway (legacy) on serving endpoints. Unity Gateway (GA August 4, 2026) governs model, model provider and MCP services as Unity Catalog securables, with service policies (Beta) as guardrails and its own budgets, inference tables and usage tables | Legacy but fully documented |
| Managed, external and custom MCP servers | Which kind fits a requirement and where its credentials live | Unity Gateway MCP services at /ai-gateway/mcp-services/<catalog>.<schema>.<name>; Genie One is system.ai.genie_one_mcp (GA). Workspace /api/2.0/mcp/... endpoints still work as legacy endpoints that bypass Unity Gateway | Same three kinds, new URLs |
| Agent Monitoring | Scoring live traffic | MLflow 3 production monitoring: scorer.register() then .start() on the experiment | Beta |
| MLflow scoring and tracing | mlflow.genai.evaluate(), built-in judges, custom scorers | Unchanged in MLflow 3. Older material uses MLflow 2 Agent Evaluation (mlflow.evaluate(model_type="databricks-agent")) | MLflow 2 forms deprecated |
Answer with the exam-guide workflow unless the question explicitly asks for the newest option. "Deploy an agent registered in Unity Catalog and get a Review App and inference tables" is agents.deploy(). "Which Agent Bricks builder answers questions about documents with citations" is Knowledge Assistant. Treat old and new names as the same product: Vector Search and AI Search, Genie space and Genie Agent, Multiagent Supervisor and Supervisor Agent.
The questions were written before these labels existed, so don't eliminate an option because today's docs call it legacy. Eliminate options that fail the requirement: guardrails or a TPM limit on an agent endpoint, Correctness on live traffic, a Production stage in Unity Catalog. The one group that is wrong outright is AI Gateway topic moderation and keyword filtering, which were removed after May 30, 2025.
How to study with this course
- Learn. Read each lesson with its self-checks. The exam callouts show how the topic is asked, and the trap callouts show the distractors.
- Lab. Run the labs on Databricks Free Edition. Steps that need a paid workspace (provisioned throughput, AI Gateway inference and usage tables, account-admin system tables) say so; read them and predict the result.
- Drill. Take the section drill untimed right after the section, and read every explanation, including the notes on the options you didn't pick.
- Flashcards. Ten minutes a day keeps the limits, defaults and names fresh: embedding windows, endpoint capacities, judge ground-truth rules.
- Mock exams. Two unseen 45-question exams with the real section counts and a 90-minute timer. Take them at the end, in one sitting each.
- Review loop. Log every miss with the rule that fixes it, re-read that lesson section, and retest the section drill.
Readiness criteria
- ≥ 80% on both mock exams on the first attempt, and no section below 70%.
- ≥ 80% on every section drill.
- You can say from memory which judges need ground truth, which AI Gateway features work on an agent endpoint, and the capacity and sync rules for both Vector Search endpoint types.
- You can name the extraction package for each source format and the chunking change that fixes each retrieval symptom.
- You can recognise each feature under its exam name and its October 2026 name.
Pacing
With up to 55 items (45 scored plus up to 10 unscored) in 90 minutes you have about 1.6 minutes per item. Most design and governance questions take under a minute; metric tables and code take longer. Use checkpoints instead of a per-question clock.
| Checkpoint | Where you should be |
|---|---|
| Minute 30 | About a third of the items answered |
| Minute 60 | About two-thirds answered |
| Minute 80 | Every item answered once, unsure ones flagged |
| Minutes 80 to 90 | Flagged items only. Change an answer only for a concrete reason |
- Answer every item on the first pass. A blank scores nothing, so always choose your best guess before you flag it.
- Flag and move on after about two minutes. Long metric tables are easier on the second pass.
- Treat every item as scored: you can't tell which ones are unscored.
Reading the questions
- Find the consumer and the cadence. Does a program or a person read the output? Does it run nightly over a table or while a user waits? That alone separates structured outputs from prose and
ai_queryfrom a serving endpoint. - Find the layer. Is the problem in the source data, chunking, retrieval, the prompt, the model, serving and access, governance or monitoring? Fix it at the layer where it starts.
- Find the deciding constraint. Latency, cost, least code, freshness, per-user permissions, no ground truth, must keep answering. The cue words below point to answers.
- Predict, then read the options. Have an answer in mind before you look, so a plausible distractor doesn't anchor you.
- Eliminate. Remove options that fix a different layer, need something the scenario lacks (ground truth, a Delta source, a supported endpoint type), or rely on a prompt instruction where a platform control exists.
| Stem wording | Usually points to |
|---|---|
| "least amount of code", "least effort" | A task AI Function, a managed index, an Agent Bricks builder or a managed MCP server, not a hand-written pipeline |
| "latency is the most critical metric" | Standard endpoint, smaller model, hybrid search and reranking off, ingestion-time masking |
| "a program consumes the output" | Structured outputs: response_format json_schema or ai_query(..., responseFormat => ...) |
| "each user may only see" | On-behalf-of-user authentication or user authorization, plus Unity Catalog row filters and masks |
| "in production", "live traffic" | Judges without ground truth, inference tables, usage tracking, production scorers |
| "promote", "roll back", "across environments" | Aliases: Prompt Registry aliases, Unity Catalog model aliases, copy_model_version, bundles |
| "must keep answering" | Mask, not Block |
Words that point to the answer
Phrases in a question that usually point to one answer. The course has the full list.
| When the question says | Think of |
|---|---|
| paragraph to a single-sentence gist, condense, shorten while keeping intent | Summarization task (ai_summarize or a chat model evaluated on summarization) |
| assign one of these labels, route each email to a department, least code | Text classification with ai_classify |
| downstream code parses the output, must always be valid JSON | Structured outputs: response_format with json_schema, or ai_query(..., responseFormat => ...) in SQL |
| expected delivery date for a transaction_id, a fact about one customer or order | A structured lookup on a table keyed by the ID (feature or online table, UC function, Genie Agent) |
| scanned images, .png or .jpeg, least code | pytesseract (OCR), or ai_parse_document on Databricks |
| tables inside PDFs must keep their rows and columns | pdfplumber (or camelot, tabula-py) |
| chunks dominated by site menus, cookie banners and footers | Remove nav, header, footer and script nodes with BeautifulSoup before extracting and chunking |
| vector database holds 100 million records but chunking produces 150 million | Larger chunks and less overlap |
| facts near the end of long chunks are never retrieved, 512-token embedding model | Truncation: re-chunk below the window or switch to databricks-gte-large-en (8,192 tokens) and re-embed |
| chunks of 512 tokens, cost and latency matter more than quality | The smallest embedding model whose context length covers 512 tokens |
Revise from the cheat sheet
The course's cheat sheet has a page for each of these:
- Design: contracts, tasks, formats, tools and Agent Bricks
- Data preparation: extraction, chunking, Delta and retrieval evaluation
- Application development: frameworks, prompts, guardrails, models and agents
- Assembling and deploying: pyfunc, registration, Vector Search, serving, MCP, prompts, CI/CD
- Governance: masking, malicious input, licensing and problematic text
- Evaluation and monitoring: judges, scorers, logs, limits, cost and SMEs
Practise with the real exam's mix
254 exam-style questions, 4 timed practice exams, flashcards and an AI tutor.