Databricks GenAI Engineer Data Preparation: 133 practice questions

12 of the 133 Data Preparation questions in the Certsqill Databricks GenAI Engineer bank, shown in full below. Each one carries an explanation for every option, not just the correct one — the wrong answers are where the marks go.

Preparing for Databricks GenAI Engineer? Take the free 5-min readiness check →

1. Retrieve child chunks and their parent sections: Which design best meets both needs?

1. Copy each parent section into every child chunk before indexing.

   Copying parent text inflates indexed content and does not cleanly separate precise retrieval from contextual presentation.
2. Index complete product sections and return matching sections.

   Complete-section retrieval preserves context but can reduce precision when a small warranty detail is buried in a long section.
3. Retrieve child chunks and their parent sections. ✓

   Child chunks support precise matching, while parent sections provide the surrounding context needed for interpretation.
4. Create equal-sized chunks without retaining document hierarchy.

   Equal-sized chunks can split related details and provide no way to restore the surrounding product context.

Retrieve child chunks with their parent sections.

2. Re-parse and clean first: What is the most likely diagnosis and remedy?

1. Re-parse and clean first, rechunk affected content, regenerate embeddings, and update the index. ✓

   The pipeline order must be restored so vectors represent the revised chunks and metadata before retrieval resumes.
2. OCR should replace parsing for all onboarding documents before another index update.

   The evidence concerns changed cleaned chunks, not image-only pages or a demonstrated OCR requirement.
3. The prompt needs more examples because changed text cannot affect vector search.

   Changed chunk text changes the indexed content and can make existing embeddings and vectors stale.
4. The index needs more overlap, regardless of whether affected embeddings were regenerated.

   Overlap may influence boundary coverage, but it cannot repair vectors generated from superseded chunks.

Changed chunks require ordered reprocessing, new embeddings, and affected index updates.

3. Run OCR, then validate serial numbers and table structure: What should it do next?

1. Discard the page because an empty result proves it has no content.

   An empty `pypdf` result can indicate an image-only page rather than an information-free page.
2. Run OCR, then validate serial numbers and table structure. ✓

   Image-only pages require OCR, and serial numbers and table structure should be checked after recognition.
3. Retry `pypdf` with a larger extraction chunk size.

   Extraction chunk size does not make `pypdf` recognize characters embedded in an image.
4. Parse the scanned page with BeautifulSoup.

   BeautifulSoup parses HTML markup; it does not recognize text contained in scanned page images.

Run OCR, then validate serial numbers and table structure.

4. Validate exceptions: What prerequisite should precede embedding and indexing?

1. Replace section-aware splitting with fixed token windows.

   Fixed windows can weaken logical context and do not validate whether cleaning preserved important qualifications.
2. Validate exceptions, units, dates, and negation. ✓

   These checks prevent cleaning from changing technical meaning before chunks become searchable.
3. Embed the cleaned chunks and inspect search results afterward.

   Post-index inspection is useful evaluation, but it comes after the required semantic-preservation checks.
4. Increase overlap between every neighboring chunk.

   Overlap may reduce boundary loss, but it does not detect meaning corrupted during boilerplate removal.

Validate retained exceptions, units, dates, and negation before indexing cleaned content.

5. Use heading-aware semantic splitting with size limits: Which TWO changes best preserve logical units while ret

Select two. More than one option is correct — every correct one is ticked below.

1. Use heading-aware semantic splitting with size limits. ✓

   Heading-aware semantic splitting keeps related material together while size limits prevent oversized retrieval units.
2. Create one chunk for each entire contract section.

   Whole sections may exceed practical context or retrieval sizes, reducing precision for focused product questions.
3. Split primarily at sentence boundaries within sections. ✓

   Sentence boundaries reduce mid-sentence cuts and preserve readable evidence inside logically grouped sections.
4. Remove headings before embedding to reduce repeated tokens.

   Removing headings discards useful context and makes section relationships harder for retrieval and answer generation.
5. Increase overlap substantially across every fixed window.

   More overlap may duplicate content without reliably keeping headings, definitions, and exceptions in one logical unit.

Combine heading-aware semantic boundaries with sentence-aware splitting inside sections.

6. Ingest and index the approved regional guide: Which change fixes the defect?

1. Increase overlap in the existing guide chunks.

   More overlap changes boundaries within existing documents but cannot introduce the missing regional guide.
2. Replace the embedding model used for guide chunks.

   A different embedding model cannot retrieve a procedure from a guide that was never processed into the index.
3. Ingest and index the approved regional guide. ✓

   Adding the authoritative guide and indexing it supplies the missing safety procedure to retrieval.
4. Increase the reranker candidate limit.

   Reranking more candidates cannot recover content absent from the indexed source collection.

Ingest and index the approved regional guide.

7. Retrieve a larger candidate pool before reranking: Which architectural change addresses the defect?

1. Tune reranker scores for rare terminology.

   Score tuning helps ordering among candidates but cannot recover documents absent from the candidate pool.
2. Replace the reranker with a larger language model.

   A stronger reranker still scores only candidates already supplied by the first retrieval stage.
3. Increase answer-generation context without changing retrieval.

   Additional context capacity cannot include a paper that initial retrieval failed to return.
4. Retrieve a larger candidate pool before reranking. ✓

   A larger first-stage pool gives the reranker an opportunity to score the missing relevant paper.

Increase first-stage candidate retrieval before applying the reranker.

8. Calculate recall@k against the truth set: What should engineers validate next?

1. Measure groundedness of generated answers.

   Groundedness assesses support from retrieved evidence, not whether retrieval found all relevant truth-set documents.
2. Calculate recall@k against the truth set. ✓

   Recall@k measures the proportion of known relevant documents retrieved within the selected cutoff.
3. Compare only average answer latency.

   Latency does not reveal how many relevant documents the retriever omitted from its results.
4. Inspect embedding vector dimensionality.

   Vector dimensionality alone does not establish whether relevant documents were retrieved for held-out queries.

Calculate recall@k using the supplied relevant-document truth set.

9. Use heading-aware semantic splitting: Which approach satisfies all three requirements?

1. Use fixed token windows with overlap.

   Fixed windows can split headings and sentences even when overlap is added.
2. Split every document into equal character ranges regardless of headings or sentences.

   Character ranges enforce a mechanical size but can fragment headings, definitions, and sentences.
3. Use one chunk per document.

   Whole-document chunks preserve structure but can exceed the required maximum size.
4. Use heading-aware semantic splitting, sentence fallback, and size limits. ✓

   This combines logical boundaries, sentence integrity, and an explicit maximum size.

Use heading-aware semantic splitting with sentence fallback and size limits.

10. Derive IDs from stable source: Which TWO controls best address the observed failure?

Select two. More than one option is correct — every correct one is ticked below.

1. Derive IDs from stable source, version, and chunk position. ✓

   Deterministic identifiers let refreshes update and delete the same logical chunks despite processing-order changes.
2. Increase chunk overlap to preserve more context.

   Overlap changes chunk content and vector volume but does not stabilize identifiers across reruns.
3. Use embedding similarity as the deletion key.

   Similarity is not a deterministic lifecycle key for addressing, updating, and deleting logical chunks.
4. Regenerate all identifiers whenever the pipeline runs.

   Regeneration deliberately creates identifier churn, making idempotent updates and deletion of prior chunks harder.
5. Compare identifiers across reruns before applying updates. ✓

   A cross-run stability check detects identifier changes even when every identifier is unique within each individual run.

Use deterministic chunk IDs and compare them across reruns before indexing.

11. Run OCR on rendered page images: What should be added?

1. Run OCR on rendered page images. ✓

   Image-only pages require OCR, such as a separately installed Tesseract-based workflow, before text can be indexed.
2. Parse the PDF with BeautifulSoup.

   BeautifulSoup parses HTML markup and does not recognize text embedded in scanned PDF page images.
3. Increase the PDF filename extension handling.

   A PDF extension identifies a container type but does not create text for image-only pages.
4. Apply ordinary embedded-text extraction again.

   Embedded-text extraction cannot recognize characters represented only as pixels in scanned pages.

Render scanned pages and run OCR to recognize their image-based text.

12. Deduplicate matching text: What should happen before embedding?

1. Add more overlap between neighboring chunks.

   More overlap repeats tokens and can create additional redundant vectors and results.
2. Increase the retrieval candidate count.

   More candidates do not remove repeated passages and can expose additional duplicates.
3. Change the embedding model for repeated paragraphs.

   A different model may change ranking but does not remove duplicate source passages.
4. Deduplicate matching text; keep source provenance. ✓

   Removing equivalent text prevents redundant vectors while retaining provenance supports governance and citations.

Deduplicate matching text while keeping provenance.

Other Databricks GenAI Engineer domains

• Application Development — 292 questions →
• Assembling and Deploying Applications — 212 questions →
• Design Applications — 136 questions →
• Evaluation and Monitoring — 116 questions →
• Governance — 78 questions →
• All 967 Databricks GenAI Engineer questions →
• Databricks GenAI Engineer certification: requirements, cost and exam format →