Databricks Certified Generative AI Engineer Associate

Learning plan

About 50 hours of focused study, split into 5 phases. Pick your weekly pace and the timeline adjusts.

Estimated finish: 5 weeks

What the exam covers

Source: official exam guide. Weights can change, so check before you start.

Before you start

• Comfortable Python and basic SQL; you will read LangChain, MLflow and PySpark snippets in questions
• A Databricks Free Edition workspace (free, no credit card) for the hands-on work
• Basic LLM vocabulary (tokens, embeddings, context window). Databricks recommends about six months of hands-on platform experience

1. Week 1 · 10 h

   Designing GenAI apps and preparing data for RAG

   • Translate a business use case into pipeline inputs/outputs and pick the model task (summarisation, extraction, classification, generation)
   • Know when to reach for Agent Bricks Knowledge Assistant, Multi-Agent Supervisor or Information Extraction instead of hand-coding a chain
   • Choose a chunking strategy (fixed, recursive, document-structure aware) based on document layout and embedding-model context length
   • Extract text from PDFs and HTML with the right Python package, strip boilerplate, and write chunks to a Delta table governed by Unity Catalog
   • Explain re-ranking and measure retrieval quality (recall@k, precision, MRR)

   Hands-on: In Free Edition, load 10–20 PDFs into a Unity Catalog volume, parse and chunk them two different ways, and write both chunk sets to Delta tables with an id, source and chunk_text column.

   Use: Exam guide (official PDF) · Agent Bricks documentation · Databricks Academy self-paced courses

2. Weeks 2–3 · 14 h

   Building chains and agents (largest domain)

   • Build a RAG chain with LangChain (or similar) that augments the prompt with retrieved context and user intent fields
   • Write prompts that force a specific output format (JSON schema, bullet list) and shift a baseline answer toward a target style
   • Select an LLM and an embedding model from model cards and experiment metrics (latency, cost, context length, quality)
   • Use MLflow and the Mosaic AI Agent Framework to log, trace and iterate on an agent; let agents query data through Genie spaces
   • Add guardrails for unsafe or off-topic input and output

   Use: Build gen AI apps on Databricks (Agent Framework) · MLflow GenAI documentation

3. Weeks 3–4 · 12 h

   Assembling, serving and governing the app

   • Create a Vector Search endpoint and Delta Sync index, query it, and size it for embedding count, update frequency, latency and cost
   • Package a chain as an MLflow pyfunc with pre/post-processing, a signature and input example, and register it to Unity Catalog
   • Serve it on Model Serving using Foundation Model APIs, control endpoint access, and use ai_query() for batch inference
   • Integrate managed, external and custom MCP servers, version prompts, and promote them across dev/test/prod
   • Apply governance controls such as PII masking, prompt-injection defences and checks on source-data licensing

   Hands-on: Build a Vector Search index over your chunk table, wrap retrieval plus a Foundation Model API call in a pyfunc chain, register it in Unity Catalog and deploy it to a serving endpoint. Then call it with ai_query() from a SQL query.

   Use: Mosaic AI Vector Search docs · Model Serving docs · MCP on Databricks · AI Functions (ai_query)

4. Weeks 4–5 · 7 h

   Evaluation and monitoring

   • Run MLflow evaluation with built-in and custom scorers, and know which LLM judges need ground truth
   • Use MLflow tracing, inference tables and Agent Monitoring to inspect a live endpoint
   • Configure AI Gateway usage tracking, inference tables and rate limits to control cost
   • Fold SME feedback from the review app back into the evaluation set

   Hands-on: Write a 20-question evaluation set for your RAG app, run mlflow.genai.evaluate with a groundedness judge and one custom scorer, change the chunk size, and compare the two runs in the MLflow UI.

   Use: AI Gateway docs

5. Week 5 · 7 h

   Practice questions and gap review

   • Work through the sample questions in the official exam guide and explain why each distractor is wrong
   • Use the AI Prep Guide prompts to generate a checklist of hands-on tasks per section, then close every gap in Free Edition
   • Re-read the Application Development and Assembling sections (52% combined) and drill the weakest objectives

   Use: Databricks AI Prep Guide (any certification) · Databricks Free Edition

You're ready to book when…

• You can build, register, serve and query a RAG chain end to end in Free Edition without following a tutorial
• You can pick a chunking strategy, embedding model and Vector Search index type for a described scenario and justify each choice
• You can name which MLflow judges need ground truth and which monitoring table answers a given production question
• You get every sample question in the exam guide right and can explain why

Exam tips

• The exam has 45 scored questions in 90 minutes and may include up to 10 unscored items you cannot identify, so answer every question carefully.
• The exam guide was refreshed in 2026 to add Agent Bricks, MCP servers and AI Gateway objectives. Practice tests written before then will miss these topics, so study from the current guide and documentation.