Study Databricks GENAI-ASSOC Model and Embedding Selection: key concepts, common traps, and exam decision cues.
Databricks expects you to choose models from evidence, not brand familiarity. The exam guide now explicitly tests model cards, context length, common experiment metrics, and embedding-model fit.
| If the requirement is mainly about… | Better first instinct |
|---|---|
| low latency and low cost with small chunks | smaller context length may fit |
| summarization or rewriting | pick the model task and then evaluate candidates |
| choosing an embedding model | consider chunk size, expected queries, and optimization strategy |
| choosing a model from a marketplace or hub | inspect model cards and metadata |
| Selection dimension | Why it matters |
|---|---|
| context length | too much context can add cost and latency without helping |
| embedding-model fit | interacts with chunking and query shape |
| model card signals | reveal task fit, limits, licensing, and operational constraints |
| experiment results | give stronger evidence than brand familiarity |
| Trap | Better rule |
|---|---|
| picking the largest context model by default | bigger context is not free and is often unnecessary |
| choosing from marketing copy alone | model cards and experiment metrics are the stronger signal |
| treating embedding choice as unrelated to chunking | they work together |
Two embedding options both work, but one is materially more expensive while the document chunks and user queries are short and narrow. What is the strongest instinct first?
Correct answer: B. Databricks expects evidence-driven selection that matches real document and query shape rather than prestige-based choice.
Model-selection questions usually reward fit rather than maximalism. Start by inspecting the model card and metadata. Then match context length, latency, and cost to the retrieval and prompt pattern. Embedding selection should be judged together with chunk design and expected user queries. The weak answer usually picks the largest model without checking the workload shape.