Study Databricks GENAI-ASSOC Pyfunc and RAG: key concepts, common traps, and exam decision cues.
The exam expects you to recognize the pieces of a deployable GenAI app, not just a notebook demo. Pyfunc packaging and clear definition of the model flavor, retriever, dependencies, signature, and examples are part of that.
| Piece | Why it matters |
|---|---|
| pyfunc model | packages logic with pre- and post-processing |
| embedding model | supports retrieval |
| retriever | supplies relevant context |
| dependencies | make the app reproducible |
| input examples and signature | clarify how the app is supposed to be used |
| If the deployment question is really about… | Better first read |
|---|---|
| what the app is supposed to accept and return | signature and examples |
| packaging reproducibility | dependencies and pyfunc structure |
| retrieval-backed answer behavior | retriever plus embedding-model setup |
| more than a raw model call | pre- and post-processing inside the pyfunc wrapper |
| Trap | Better rule |
|---|---|
| treating deployment like “just run the notebook” | package the behavior intentionally |
| ignoring pre- and post-processing | pyfunc often wraps more than a bare model call |
| forgetting signature or examples | contract clarity matters in deployment |
A Databricks team can run a RAG notebook interactively, but no one can tell what inputs the deployed app expects or which preprocessing steps it applies. What is the strongest first fix?
Correct answer: A. GENAI-ASSOC deployment questions reward intentional packaging and interface clarity, not notebook-only success.
Deployment questions usually start with the deployable contract. If the issue is packaging chain logic for deployment, think pyfunc. If the issue is how the RAG application is structured, separate model flavor, embeddings, retriever, dependencies, and signature. The weak answer usually treats deployment as just “serve a model” without clarifying the application contract.