Databricks GenAI Engineer cheat sheet
Last-minute reference for the Databricks Generative AI Engineer Associate exam.
Exam facts
| Questions | 45 scored |
| Duration | 90 minutes |
| Fee | 200 USD |
| Validity | 2 years |
| Passing score | Not published by Databricks |
Section weights
| Section | Weight |
|---|---|
| Application Development | 30% |
| Assembling and Deploying Apps | 22% |
| Design Applications | 14% |
| Data Preparation | 14% |
| Evaluation and Monitoring | 12% |
| Governance | 8% |
Development plus deployment is 52%.
RAG quick answers
| Symptom or requirement | Answer |
|---|---|
| 40,000 docs updated weekly | RAG over an index refreshed as docs change |
| Answers must cite the source paragraph | Retrieval returns identifiers and locations |
| Topically related but useless passages | Smaller chunks with overlap |
| Tables lose meaning when flattened | Preserve row/column relationships |
| Exact identifiers missed by semantic search | Hybrid search with keyword matching |
| Output must be fixed JSON | Constrain to a schema and validate |
| Latency dominated by 50 chunks, quality flat above 8 | Retrieve fewer, rerank to keep the best |
| Hidden instructions in retrieved content | Isolate retrieved content from the instruction channel |
| Tool charges a customer | Explicit confirmation before execution |
Deployment
- Tenfold month-end spike → autoscaling with min and max
- Compare two prompt versions on live traffic → split traffic, compare measured outcomes
- Reproducible six months later → version the full configuration: prompt, model version, retrieval settings, code
- Must not be publicly reachable → private network access, not an IP allow-list
Governance and evaluation
- Entitlements enforced at retrieval, before content enters the context
- Log metadata, source references and decisions, sensitive content redacted
- Groundedness = supported by sources; relevance = addresses the question
- Quality decline with no code change → drift in the question distribution or the corpus
The published-score trap
No passing score is published. Calibrate on a weighted practice test.
Night-before checklist
- Six sections; development plus deployment is 52%
- Chunking failure mode and its fix
- Groundedness versus relevance
- Check ID and proctoring rules — see exam day
Take the 20-question practice test.