Databricks Certified Generative AI Engineer Associate tests whether you can build, deploy and run LLM and agent applications on Databricks. It covers prompt and chain design, RAG data preparation and chunking, Mosaic AI Vector Search, MLflow and the Agent Framework, model serving with Foundation Model APIs, MCP servers, guardrails, and evaluation and monitoring with AI Gateway and inference tables. The current exam was published on 18 March 2026.
GenAI engineer and AI engineer openings increasingly ask for Databricks along with LangChain and RAG skills. This guide follows the official exam guide: six sections with weights, the Databricks pieces of a RAG agent, a 6-week plan and 15 original scenario questions.
Quick facts (Databricks, November 2026)
| Item | Detail |
|---|---|
| Questions | 45 scored (multiple choice and multiple selection), plus up to 10 unscored items |
| Time | 90 minutes (extra time for unscored items is included) |
| Fee | USD 200 plus applicable taxes |
| Languages | English, Japanese, Brazilian Portuguese, Korean |
| Delivery | Online-proctored or at a test centre |
| Prerequisites | None; 6 months of hands-on Databricks experience recommended |
| Validity | 2 years; recertify by taking the current exam |
Read the official Generative AI Engineer Associate page and download the current exam guide before booking.
Sections and weights
| Section | Weight |
|---|---|
| 1. Design Applications | 14% |
| 2. Data Preparation | 14% |
| 3. Application Development | 30% |
| 4. Assembling and Deploying Applications | 22% |
| 5. Governance | 8% |
| 6. Evaluation and Monitoring | 12% |
What each section covers
- Design (14%): prompts that return a specific format; choosing model tasks; chain components; turning business goals into pipeline inputs and outputs; ordering tools for multi-stage reasoning; when to use Agent Bricks (Knowledge Assistant, Multi-Agent Supervisor, Information Extraction).
- Data preparation (14%): chunking strategies for document structure and model limits; filtering noisy content; Python packages for document extraction; writing chunks to Delta tables in Unity Catalog; picking source documents; retrieval metrics; advanced chunking; re-ranking.
- Application development (30%): LangChain and similar tools; spotting quality and safety issues; prompt augmentation; guardrails; choosing LLMs and embedding context length; model cards and marketplaces; experiment metrics; MLflow and Agent Framework for agents; evaluation vs monitoring; multi-agent systems that use Genie Spaces or the conversational API.
- Assembling and deploying (22%): pyfunc chains with pre- and post-processing; access control from serving endpoints; RAG building blocks (model flavour, embedding model, retriever, signature); registering models in Unity Catalog; creating and querying a Vector Search index; Foundation Model APIs; batch inference with
ai_query(); memory stores; CI/CD for agents and prompts; managed, external and custom MCP servers; prompt versioning; user interfaces (Apps, Slack, Teams). - Governance (8%): masking as a guardrail; protecting against malicious inputs; licensing of data sources; mitigating problematic text.
- Evaluation and monitoring (12%): choosing models from metrics; deployment metrics; MLflow scoring and tracing; inference logging; cost control; inference tables and Agent Monitoring; judges that need ground truth; AI Gateway (inference tables, usage tables, rate limits); custom scorers; SME feedback.
A RAG agent on Databricks, piece by piece
| Step | Databricks component |
|---|---|
| Store and govern source documents | Unity Catalog volumes and Delta tables |
| Parse and chunk documents | Notebooks or jobs writing chunks to a Delta table |
| Embed and index | Mosaic AI Vector Search (Delta Sync index) |
| Call an LLM | Foundation Model APIs or external models through Model Serving |
| Build and log the agent | MLflow with the Agent Framework; register in Unity Catalog |
| Serve | Model Serving endpoint; Databricks Apps for a UI |
| Evaluate and monitor | MLflow evaluation and tracing, inference tables, AI Gateway |
# Query a Vector Search index from a notebook (databricks-vectorsearch)
from databricks.vector_search.client import VectorSearchClient
vsc = VectorSearchClient()
index = vsc.get_index(endpoint_name="vs-endpoint", index_name="main.rag.policy_chunks_index")
hits = index.similarity_search(
query_text="How many casual leaves do new joiners get?",
columns=["chunk_id", "chunk_text", "source"],
num_results=5,
)
-- Batch inference over a table with ai_query() in SQL
SELECT ticket_id,
ai_query('databricks-meta-llama-3-3-70b-instruct',
CONCAT('Classify this support ticket as billing, bug or other: ', body)) AS label
FROM main.support.tickets;Endpoint and model names change often. Use the names listed in your workspace's Serving page.
6-week study plan
| Week | Focus | Lab |
|---|---|---|
| 1 | Prompt design, model tasks, chains, Agent Bricks | Prompt that returns valid JSON every time |
| 2 | Document parsing, chunking, Delta tables, re-ranking | Chunk 50 PDFs into a Unity Catalog table |
| 3 | Vector Search, embeddings, Foundation Model APIs, LangChain | Working RAG chain over your chunks |
| 4 | MLflow Agent Framework, pyfunc, Unity Catalog registration, Model Serving, MCP servers | Deploy the agent to a serving endpoint |
| 5 | Guardrails, masking, licensing, ai_query batch jobs, CI/CD for prompts | Add an input guardrail and a batch classification job |
| 6 | MLflow evaluation and tracing, judges, inference tables, AI Gateway, cost | Evaluation set plus rate limits; book the exam |
15 scenario questions with answers
These are original scenarios written for this guide, not exam questions.
1. Answers must always come back as JSON with three fixed keys. What do you change first?
The prompt: specify the exact format, with an example of the output.
2. Long HR policy PDFs have clear headings. Which chunking strategy fits?
Structure-aware chunking by section or heading, with some overlap.
3. Retrieval returns relevant chunks, but the best one is often ranked fifth. What helps?
A re-ranking step after the initial retrieval.
4. Where should chunked text live so it is governed and can sync to Vector Search?
A Delta table in Unity Catalog.
5. The index must update automatically when source rows change. Which index type?
A Delta Sync index in Mosaic AI Vector Search.
6. A million product reviews need sentiment labels overnight. What is the simplest approach?
Batch inference with ai_query() in SQL.
7. A chain needs custom pre-processing before the LLM and post-processing after. How do you package it?
As an MLflow pyfunc model.
8. The agent must answer data questions over sales tables in natural language. What can it call?
A Genie Space (or the conversational API).
9. The agent needs tools from an external system through a standard protocol. What do you integrate?
An MCP server (managed, external or custom).
10. Users try to make the bot reveal internal instructions. What do you add?
Input guardrails that detect and block prompt-injection attempts.
11. Customer phone numbers must not reach the LLM. What technique?
Masking personal data before it is sent to the model.
12. You need to see each step of an agent's reasoning and tool calls. What do you use?
MLflow tracing.
13. Which kind of evaluation judge needs ground-truth answers?
Correctness judges that compare against expected answers; safety and relevance judges usually don't.
14. A team's agent is running up the LLM bill. Where do you add limits and track usage?
AI Gateway rate limits and usage tables on the serving endpoint.
15. Subject experts disagree with some answers. How do you use their input?
Collect their feedback with the review app, add it to the evaluation set and re-evaluate after changes.
After the certification
- Practise GenAI engineer interview questions, LangChain interview questions and agentic AI interview questions.
- Read FDE vs AI engineer vs GenAI engineer and GenAI engineer salaries in India, and pair this with the Databricks Data Engineer Associate.
- See live AI and GenAI jobs.
Cloud Soft Solutions in Ameerpet, Hyderabad teaches generative AI, RAG, agents, Databricks and MLOps with real-time projects. See the APEX AI, ML, Cloud and Cyber Security program and all courses.
📞 Call +91 96660 19191 · WhatsApp · Flat No. 513, 5th Floor, Nilagiri Block, Aditya Enclave, Ameerpet, Hyderabad
Frequently Asked Questions
How many questions are on the Databricks Generative AI Engineer Associate exam?
45 scored multiple-choice and multiple-selection questions, plus up to 10 unscored items, in 90 minutes.
How much does the exam cost?
USD 200, plus applicable taxes.
Which section has the most weight?
Application Development, at 30%, followed by Assembling and Deploying Applications at 22%.
When was the current exam published?
18 March 2026.
Does the exam cover agents and MCP?
Yes. It covers the MLflow Agent Framework, Agent Bricks, multi-agent systems with Genie Spaces, and integrating managed, external and custom MCP servers.
How long is the certification valid?
2 years. You recertify by taking the current exam.
