Databricks GenAI Assoc Tests RAG and Agent Plumbing on Databricks — Not Prompt Tricks
For engineers building retrieval and agent apps with Vector Search, MLflow, Model Serving and Agent Bricks. The shift: let the stated constraint (latency, cost, freshness, ground truth) pick the component, not the most powerful option.
Check Your Readiness →Most candidates understand Databricks Certified Generative AI Engineer Associate concepts — and still fail. This exam tests how you apply knowledge under pressure.
The current guide covers the version live since 18 March 2026. Questions mix short business scenarios with code fragments such as MLflow logging, Vector Search client calls or ai_query in SQL. Each stem names one deciding attribute, like vector count, update frequency, latency or whether expected answers exist.
Increase top-k so more chunks reach the prompt
Recall is fine; ranking is the weakness. Add a re-ranking stage to reorder candidates by relevance. A reranker cannot help when the right chunk is absent, and a larger top-k only adds noise and tokens.
Deploy Agent Bricks Knowledge Assistant over the contracts
Knowledge Assistant returns cited chat answers. Agent Bricks Information Extraction is the managed template that turns unstructured documents into structured fields, which is exactly the rows-and-columns output the scenario asks for.
Standard endpoint with continuous sync
Pick a storage-optimized Vector Search endpoint, built for very large vector counts at lower cost with somewhat higher latency, and use triggered sync after each weekly refresh. Continuous sync runs constantly and costs more.
Rewrite the system prompt to be more careful with numbers
Enable MLflow Tracing to record each tool call, retrieval and LLM step with inputs and outputs. Traces show whether the order lookup, the retriever or the final generation produced the bad value before anything is changed.
Delta Sync Vector Search indexes track a source Delta table, so the chunk table needs a primary key and delta.enableChangeDataFeed set. Direct Vector Access indexes never sync from tables; they are written through the API, which suits embeddings produced outside Delta.
Correctness and retrieval sufficiency compare output with ground truth. Groundedness, relevance to query, retrieval relevance and safety are label-free, so those are the judges that work in production monitoring, where live requests never carry expected facts.
Models registered under a catalog.schema.model name have no Staging or Production stages. Assign an alias such as champion and load with models:/name@champion. Registration also needs a signature and input example, and a bare two-level name fails.
When millions of rows need inference and nobody waits on the result, ai_query() in SQL runs it in parallel against a serving endpoint and writes back to a table. Real-time endpoints called from Python loops are meant for interactive traffic.
Inference tables store request and response payloads for debugging and evaluation. Token and request consumption by caller lives in the AI Gateway usage tracking system tables, which are the source for chargeback; rate limits then cap spend per user or endpoint.
Strengthen weak areas with exam-style practice questions and detailed explanations.
Simulate the real exam experience and assess your readiness under timed conditions.
Review key concepts, objectives, and exam topics in one place.
Get personalized explanations, learning recommendations, and instant answers.
Follow a structured learning path designed to help you prepare efficiently.
Check whether you can match Vector Search, serving and judges to each stated constraint.