Databricks Generative AI Engineer Associate

The exam

45 scored questions (+ up to 10 unscored) · 90 min · exam published 18 March 2026.

Weights

New names

Vector Search → Databricks AI Search · Genie spaces → Genie Agents · endpoint AI Gateway (legacy) → Unity Gateway · Knowledge Assistant, Supervisor Agent, AI Playground, agents on Model Serving (agents.deploy) now listed as legacy → Genie Agents (low code), Agent Bricks CLI on Agent Runtime (code) · Asset Bundles → Declarative Automation Bundles. Questions may use either name.

Goal → task

Chain = fixed steps; agent = LLM picks the next tool. Strict format: rules in the system prompt + an example output; response_format JSON schema when code parses it.

Agent Bricks

Tools: one job each, clear description; gather first, then act. Evaluation-driven: metrics and eval set before building.

Parse and clean

Strip menus, headers, footers, cookie banners; de-duplicate chunks. Missing answers (shipping dates) → add the right source (a table keyed by ID), not more prompt.

Chunking

Fixed + overlap (baseline) · recursive (prose) · structure-aware (headings) · semantic · small-to-big (parent document). Chunk ≤ embedding model's context. Fewer records: bigger chunks, less overlap. Keep metadata per chunk.

Write: chunk → row + primary key → UC Delta table → Change Data Feed on → Delta Sync index.

Indexes and retrieval

Retrieval metrics: precision@k, recall@k, MRR, NDCG. Re-ranker reorders candidates with a more accurate model; it can't find what retrieval missed.

Prompts and parameters

System prompt = role, rules, format. Few-shot fixes format; chain-of-thought for multi-step logic. temperature, top_p = randomness; max_tokens caps output; context window = input + output. Hallucinations: RAG, allow "I don't know", cite sources.

Choosing models

Foundation Model APIs

Mostly OpenAI-compatible; stream=True streams tokens.

Frameworks and tools

databricks-langchain: ChatDatabricks, DatabricksVectorSearch (.as_retriever()), UCFunctionToolkit (function_names=[…], then .tools), GenieAgent. LCEL: prompt | llm | parser. UC functions as tools = governed (grants, audit); COMMENTs describe them.

Agents, guardrails, Genie

Author with ResponsesAgent (ChatAgent is older). Guardrails: input (safety model, injection detection), output (safety, PII masking), AI Gateway safety + PII on model endpoints (Public Preview). Table questions → Genie space (Conversation API, Genie MCP, GenieAgent). Evaluation = offline dataset; monitoring = live traffic.

Package and deploy

Pyfunc: load_context (clients, once) + predict with pre- and post-processing. Models from code: set_model(AGENT), then python_model="agent.py". Log resources, pip_requirements, input_example (signature = schema). Register catalog.schema.model; promote with aliases. agents.deploy(name, version) → endpoint, review app, tracing.

Access and querying

Query: DatabricksOpenAI(), then .responses.create(model, input); ChatAgent → chat.completions with messages. Idle, variable traffic → scale to zero (not for production).

Vector Search

Endpoint STANDARD or STORAGE_OPTIMIZED · pipeline_type TRIGGERED (idx.sync()) or CONTINUOUS · similarity_search(query_text, columns, num_results, filters, query_type="HYBRID").

Batch, memory, MCP

Millions of rows on a schedule → ai_query or ai_classify, ai_extract. Memory → Lakebase (CheckpointSaver, DatabricksStore) or Delta tables. MCP: managed (Genie, AI Search, UC functions, DBSQL) → least effort; external → UC HTTP connection holds credentials; custom → Databricks Apps.

Prompts, CI/CD, UI

Prompt Registry (Beta): register_prompt (UC name, {{var}}), set_prompt_alias, load_prompt("prompts:/name@alias"); versions immutable, aliases move. CI/CD: bundles, test each tool, evaluate as a gate, promote by alias, sync the index. UI: Databricks Apps (backend calls the endpoint as the app or as the user); never tokens in the browser.

Data first

Evaluate

mlflow.genai.evaluate(data=, predict_fn=, scorers=): records with inputs (and expectations). Tracing (autolog, @mlflow.trace): spans show which step failed.

Custom: @scorer on inputs, outputs, expectations, trace → bool, number, string or Feedback; make_judge for LLM criteria. Inconsistent SME ratings → rubrics + calibration.

Monitor and control cost

Cost: smaller model, shorter prompts, right serving mode, scale to zero, batch ai_query. Size: smallest model that meets the target on the same eval set.