Design Applications
14%of exam
Prompt designTask selectionChain componentsBusiness mappingSolution decomposition
Data Preparation
14%of exam
ChunkingEmbeddingsRetrieval qualityDocument extractionDelta LakeUnity Catalog
Application Development
30%of exam
RAG chainsAgentsTool callingLangChainLangGraphGuardrails
Assembling + Deploying
22%of exam
MLflowModel ServingVector SearchPyfuncCI/CDFoundation Model APIs
Governance
8%of exam
MaskingGuardrailsLicensingUnity CatalogAI GatewaySafe input
Evaluation + Monitoring
12%of exam
Offline evalOnline monitoringLLM judgeInference tablesRAGASCost control
Quick Facts
- Exam
- Databricks GenAI
- Credential
- GenAI Engineer Associate
- Time
- 90 min
- Pass
- 70% (32/45)
- Questions
- 45 scored
- Fee
- $200
- Blueprint
- Mar 18 2026
Design Decisions
- Problem decomposition
- Split to subtasks
- Task selection
- Pick model type
- Chain design
- Sequence steps
- Business mapping
- Goal to behavior
- Architecture choice
- Pick pattern
- Cost/latency
- Tradeoff balance
Prompt Engineering
- Zero-shot
- No examples
- Few-shot
- Example-driven
- Chain-of-thought
- Step reasoning
- System prompt
- Set behavior
- Role prompt
- Persona
- Template
- Reusable scaffold
- Structured output
- JSON/schema
Semantic vs Keyword Search
Semantic
- Vector similarity
- Meaning match
- Needs embeddings
Keyword
- Term match
- BM25/SQL
- No embeddings
Meaning vs terms
Chunking Strategy Picker
- Uniform docs→Fixed-size(N tokens)
- Sentences matter→Sentence(Split by .)
- Topic shifts→Semantic(Meaning bounds)
- Nested structure→Recursive(Hierarchy)
- Short docs→Whole doc(No split)
- Need context→Overlap(Share text)
- Retrieve precision→Parent-child(Child index)
Chunking Strategies
- Fixed-size
- N tokens each
- Sentence
- Split by sentence
- Semantic
- Topic boundaries
- Recursive
- Hierarchical split
- Overlap
- Adjacent share text
- Parent-child
- Index child; retrieve parent
- Sliding window
- Fixed stride
Data Preparation Pipeline
- Extraction
- Parse documents
- Filtering
- Drop noise
- Cleaning
- Normalize text
- Delta Lake
- Versioned tables
- Unity Catalog
- Data governance
- Indexing
- Build vector index
RAG Pipeline
Ingest | Chunk | Embed | Index | Retrieve | Rerank | Generate
Ingest: parseChunk: splitEmbed: vectorIndex: storeRetrieve: fetchRerank: reorderGenerate: answer
RAG vs Fine-tuning
RAG
- Inject knowledge
- No training
- Updates easy
Fine-tuning
- Bake behavior
- Needs training
- Updates costly
Knowledge vs behavior
RAG vs Fine-tune vs Prompt
- Need latest info→RAG(Inject knowledge)
- Need private data→RAG(Vector Search)
- Need new skills→Fine-tune(Bake behavior)
- Need behavior shift→Fine-tune(LoRA/QLoRA)
- Need format only→Prompt(No training)
- Quick one-shot task→Prompt(Zero-shot)
- Multi-step task→Agent(Tool calling)
- Need external tools→Agent(MCP server)
RAG Components
- Vector store
- Semantic index
- Embedding
- Vector representation
- Chunking
- Split documents
- Retrieval
- Fetch relevant chunks
- Re-ranking
- Reorder top results
- Context window
- Max input tokens
- Grounding
- Cite source facts
- Augmentation
- Inject context
LLM Params
Temp=creativity | Top-p=nucleus | Top-k=pool
Temp: randomTop-p: nucleusTop-k: candidates
LoRA vs Full Fine-tune
LoRA
- Low-rank adapters
- Few params
- Cheap train
Full
- All weights
- Costly train
- Max control
Adapters vs all weights
LLM Concepts
- Token
- Subword unit
- Context window
- Max input tokens
- Temperature
- Randomness control
- Top-p
- Nucleus sampling
- Top-k
- Top candidates
- Hallucination
- Unfounded output
- Sampling
- Probabilistic decode
Agent Bricks vs Framework
Agent Bricks
- Managed
- Low code
- Databricks-hosted
Framework
- Custom code
- LangChain
- Full control
Managed vs custom
Fine-tuning Methods
- Full fine-tune
- Update all weights
- LoRA
- Low-rank adapters
- QLoRA
- Quantized LoRA
- Instruction-tune
- Follow directions
- PEFT
- Partial tuning
- RLHF
- Human feedback
- DPO
- Preference optimize
Agent Patterns
- ReAct
- Reason + act
- Tool calling
- Invoke functions
- Planning
- Decompose tasks
- Reflection
- Self-critique
- Multi-agent
- Coordinate roles
- MCP server
- Tool protocol
- Memory
- Persist context
Deploy Gates
Package | Register | Serve | Monitor
Package: pyfuncRegister: MLflowServe: endpointMonitor: drift
Playground vs Model Serving
Playground
- Interactive test
- Prompt iterate
- No endpoint
Model Serving
- Deploy endpoint
- Production API
- Scale traffic
Test vs deploy
Deployment Picker
- Track experiment→MLflow(Log params)
- Register model→MLflow register(Version)
- Real-time API→Model Serving(Endpoint)
- Async scoring→Batch inference(No endpoint)
- Semantic search→Vector Search(Index)
- Usage metering→AI Gateway(Track)
- Managed agents→Agent Bricks(Low code)
Databricks Tools
- Vector Search
- Semantic index
- AI Playground
- Test prompts
- Foundation Model APIs
- Hosted base models
- MLflow
- Track experiments
- Unity Catalog
- Governance hub
- Model Serving
- Deploy endpoints
- AI Gateway
- Usage tracking
- Agent Bricks
- Managed agents
MLflow vs Unity Catalog
MLflow
- Track experiments
- Log params
- Version models
Unity Catalog
- Govern data
- Access control
- Lineage
Track vs govern
Deployment Patterns
- Pyfunc
- Python packaging
- MLflow register
- Version models
- Model Serving
- Real-time endpoint
- Batch inference
- Async scoring
- CI/CD
- Automate release
- Persistent state
- Save context
- Streaming
- Real-time input
Governance Stack
UC | Masking | Guardrails | AI Gateway
UC: governMasking: PIIGuardrails: filterGateway: track
Governance Controls
- Masking
- Hide PII
- Guardrails
- Input/output filters
- Licensing
- Model terms
- Unity Catalog
- Central governance
- AI Gateway
- Track usage
- PII detection
- Find sensitive
- Audit logging
- Track actions
Eval Dimensions
Faithfulness | Relevance | Context | Answer
Faithfulness: groundedRelevance: matchedContext: recallAnswer: quality
Offline vs Online Eval
Offline
- Pre-deploy
- Ground truth
- RAGAS/BLEU
Online
- Production
- Live metrics
- Drift detect
Test vs monitor
Evaluation Metrics
- BLEU
- N-gram overlap
- ROUGE
- Recall overlap
- RAGAS
- RAG eval suite
- Human eval
- Gold standard
- LLM judge
- Auto evaluate
- Faithfulness
- Source-grounded
- Groundedness
- No hallucination
Common Traps
RAG vs Fine-tune vs Prompt
RAG adds knowledge ≠ Fine-tune shifts behavior
Temperature vs Top-p
Temp scales randomness ≠ Top-p truncates nucleus
Vector vs Keyword Search
Semantic matches meaning ≠ Keyword matches terms
Offline vs Online Eval
Offline needs ground truth ≠ Online detects drift
MLflow vs Model Serving
MLflow tracks experiments ≠ Serving deploys endpoints
Unity Catalog vs Delta Lake
UC governs access ≠ Delta stores tables
Agent Bricks vs Framework
Bricks is managed ≠ Framework is custom code
Last Minute
- 1.14/14/30/22/8/12
- 2.45 Q / 90 min / $200
- 3.Pass: 70% (32/45)
- 4.RAG = retrieve + augment
- 5.LoRA = adapters; QLoRA = quantized
- 6.Vector Search = semantic index
- 7.MLflow = track; Serving = deploy
- 8.Unity Catalog = governance
- 9.AI Gateway = usage tracking
- 10.Agent Bricks = managed agents
- 11.MCP = tool protocol
- 12.March 18 2026 = blueprint update
- 13.14-day retake; full fee
Buy on AmazonSame family resources
Explore More Databricks Certifications
Continue into nearby exams from the same family. Each card keeps practice questions, study guides, flashcards, videos, and articles in one place.
More From This Family
Videos and articles for deeper review.
VideoDatabricks Certified Data Analyst Associate Exam Guide 2026: SQL, AI/BI Dashboards, Genie, and Free PracticeVideoDatabricks Data Engineer Associate Exam Guide (2026): Pass on Your First TryArticleDatabricks Certified Data Analyst Associate Exam Guide 2026: SQL, AI/BI Dashboards, Genie, and Free Practice14 min readArticleDatabricks Data Engineer Associate Exam Guide (2026): Pass on Your First Try26 min read