Databricks GenAI design and data preparation
Design Applications and Data Preparation are 14% each — 28% of the Databricks Generative AI Engineer Associate exam between them, roughly 12 of the 45 scored questions. They cover the decisions made before any application code exists, and the data work that determines whether it will ever work well.
Designing a generative AI application
The first decision is what approach the requirement calls for:
| Approach | Use when |
|---|---|
| Prompting alone | The model already knows enough; you need format or tone control |
| RAG | Answers must come from your content, especially content that changes |
| Fine-tuning | You need consistent behaviour, style or domain vocabulary |
| Chains and agents | The task needs multiple steps or tool use |
The rule that decides most exam questions: if the information changes, it is RAG. If the behaviour needs to change, it is fine-tuning. A scenario mentioning documents updated weekly, current data, or citing sources is a RAG question every time.
Alongside that sits model selection — capability against cost against latency — and designing the components: source data, chunking, embeddings, index, retriever, prompt, model, output handling.
Data preparation: the part that decides quality
Most RAG failures are retrieval failures, and most retrieval failures are data preparation failures. This section is small in weighting and large in consequence.
Chunking is the central decision:
- Too large and each chunk covers several topics, so the embedding represents none of them well and retrieval becomes imprecise.
- Too small and chunks lose the context needed to make sense on their own.
- Overlap between chunks prevents a relevant passage being split down the middle and lost.
- Chunking that respects document structure — sections, headings, paragraphs — generally beats fixed character counts.
Embeddings turn each chunk into a numeric representation of meaning, so semantically similar content can be found even when the wording differs. The embedding model used for indexing must be the same one used for queries.
Extraction and cleaning comes first: getting usable text out of PDFs, HTML and documents, removing boilerplate, and preserving structure like tables and headings where it carries meaning.
Index maintenance is ongoing. A corpus that changes needs a refresh strategy, and stale indexes are a common production problem.
The diagnostic worth memorising
When a RAG application gives a wrong answer:
- Was the right content retrieved? If not, the problem is data preparation or retrieval — chunking, embeddings, index freshness.
- Was the right content retrieved but the answer still wrong? Then look at the prompt or the model.
Step one is the answer far more often, and the exam reflects that.
Sample questions
Question 1. A team indexes long policy documents as single chunks of several thousand words each. Retrieval frequently returns documents that are only loosely relevant. What should change first?
- A. Split documents into smaller overlapping chunks aligned to sections
- B. Switch to a model with more parameters
- C. Return twenty results instead of five
- D. Increase the temperature to encourage broader answers
Show answer
Answer: A
Very large chunks produce embeddings that average across many topics, so matching becomes imprecise; splitting into smaller overlapping chunks aligned to document structure directly addresses this. A larger model does not fix what is retrieved, more results returns more loosely relevant content, and raising temperature affects generation rather than retrieval.
Question 2. An application must answer questions about a product catalogue that changes weekly, and must reflect changes within a day. Which design fits?
- A. Fine-tune the model weekly on the catalogue
- B. Prompt engineering alone with detailed instructions
- C. RAG with a scheduled index refresh
- D. A model with a larger context window
Show answer
Answer: C
RAG over an index refreshed on a schedule keeps answers current without touching the model, which is what a weekly-changing catalogue requires. Fine-tuning weekly is expensive and repeats indefinitely, prompting alone gives the model no access to the catalogue, and a larger context window does not supply the data.
Question 3. A team embeds its document corpus with one embedding model, then switches the query path to a different embedding model. What is the consequence?
- A. Retrieval becomes slightly slower but remains accurate
- B. Retrieval quality collapses, because queries and content are no longer in the same vector space
- C. Storage costs increase but results are unaffected
- D. Nothing changes, since embeddings are interchangeable
Show answer
Answer: B
Queries and stored content must be embedded by the same model to occupy the same vector space, so mixing models makes similarity comparisons meaningless and retrieval quality collapses. Costs and latency may change slightly, but the decisive effect is that retrieval stops working correctly.
What to practise
Index the same document set three ways — one large chunk per document, small chunks with no overlap, and structure-aware chunks with overlap — and run the same five questions against each. The difference in answer quality with the model held constant is the clearest possible demonstration of why this section exists.