Databricks Generative AI Engineer Associate Course Notes

Design the application before choosing products: define prompt/output contracts, decompose tasks, select only necessary chain components, translate business goals into explicit inputs/outputs and action boundaries, order tools by dependency and risk, and choose the simplest Agent Bricks pattern that fits.

• Choosing architecture before defining safe actions.

• Read the requirement nouns first: input, output, source, action, latency, cost, authorization, or evaluation evidence.

RAG quality starts before retrieval. Choose authoritative sources, extract them correctly, remove boilerplate, preserve semantic structure during chunking, stay within embedding limits, store stable chunk lineage in Delta, and measure retrieval at the same cutoff the generator receives.

• Trying a larger generator context after the embedding model already truncated evidence.

• Read the requirement nouns first: input, output, source, action, latency, cost, authorization, or evaluation evidence.

Build and iterate using explicit task requirements and measurement: choose frameworks for orchestration needs, diagnose response-quality dimensions, tune chunking/prompts from experiments, layer guardrails, select models by hard capability/governance constraints first, and instrument agents with MLflow tracing and evaluation.

• Choosing the cheapest/newest model before verifying required capabilities.

• Read the requirement nouns first: input, output, source, action, latency, cost, authorization, or evaluation evidence.

Package, register, serve, govern, and release reproducibly. Use pyfunc for deterministic Python boundaries, separate endpoint access from downstream resource permissions, know AI Search/Vector Search patterns, choose batch versus interactive inference correctly, persist durable agent state when needed, and promote immutable model/prompt/index components through tested CI/CD.

• Treating a missing runtime package as an authorization problem.

• Read the requirement nouns first: input, output, source, action, latency, cost, authorization, or evaluation evidence.

Govern the data path and action path. Minimize/mask sensitive data, treat user and retrieved text as untrusted, enforce least-privilege tools, track source rights/provenance, and repair or exclude problematic source text at the pipeline layer.

• Masking only the response while storing raw unnecessary PII in logs.

• Read the requirement nouns first: input, output, source, action, latency, cost, authorization, or evaluation evidence.

Measure what matters at the right stage. Apply hard quantitative SLOs before optimization, trace and score agents at intermediate steps, log governed production evidence, monitor semantic quality plus operations/cost/tool behavior, and turn validated SME/user feedback into durable evaluation cases.

• Selecting by one metric after another hard requirement fails.

• Read the requirement nouns first: input, output, source, action, latency, cost, authorization, or evaluation evidence.