Databricks-GenAI-Assoc Data Preparation Practice Question

Answer analysis

Option-by-option breakdown

For each option: why learners choose it and why it is or isn't the right answer here.

• ✗

  Remove stop words and punctuation from every passage before generating embeddings.

  Why it's wrong here

  Modern embedding models rely on full natural-language context, including function words and punctuation, to place passages in vector space. Stripping them can distort meaning and hurt retrieval, and because tokenizers compress common words efficiently, the token savings are small compared with the semantic damage, making this a poor trade for embedding quality.

• ✓

  Deduplicate near-identical passages and drop documents that fail a minimum content-length threshold.

  Why this is correct

  Near-duplicate passages waste embedding calls and crowd the index with redundant vectors, so removing them lowers cost and improves result diversity. Dropping documents below a minimum length eliminates fragments and stubs that produce low-information embeddings, which raises overall retrieval precision without discarding meaningful content.

• ✗

  Truncate every document to its first 512 characters to guarantee uniform chunk size.

  Why it's wrong here

  Blind truncation discards the majority of long documents, so retrieval can never surface information beyond the cut point. It also produces chunks that end mid-sentence, harming embedding coherence. Uniform size is better achieved by semantic chunking with overlap, not by cutting content away from the source.

• ✓

  Normalize whitespace and strip boilerplate such as navigation menus, headers, and repeated legal footers.

  Why this is correct

  Boilerplate and irregular whitespace inflate token counts without carrying meaning, so removing them lowers embedding cost and sharpens the semantic signal in each chunk. Normalization also reduces spurious variation between otherwise identical passages, which improves similarity matching during retrieval and keeps chunk boundaries aligned with real content.

• ✗

  Convert all text to uppercase before tokenization to standardize casing.

  Why it's wrong here

  Uppercasing destroys casing cues that tokenizers and models use to distinguish proper nouns, acronyms, and sentence structure, which can degrade embedding quality. It does not reduce token count meaningfully because most tokenizers already handle case variants, so it adds preprocessing risk without the cost or retrieval benefit the team needs.