4.2 Preparing Data for AI and ML PROFESSIONAL-DATA-ENGINEER Practice Questions (Page 4)

Questions 16–20

1. 16foundation · easy

   In the context of retrieval-augmented generation (RAG), what is the purpose of chunking documents before generating embeddings?

   • ATo convert documents into a format that can be processed by a language model.
   • BTo encrypt the documents before storing them.
   • CTo improve the relevance of retrieved passages by breaking documents into smaller, focused pieces.
   • DTo reduce the size of the documents so they can be stored in a vector store.
   Reveal Solution Discussion Select an answer first
2. 17expert · hard

   A telecom company is developing a churn prediction model using BigQuery ML. The dataset includes a `call_duration` feature with many outliers, a `plan_type` categorical feature, and a `customer_tenure` feature. The data engineer needs to prepare the data for both training and batch prediction. They also need to ensure that the model is not overly influenced by extreme values. Which approach best addresses these requirements?

   • AUse min-max scaling on `call_duration`, drop `plan_type`, and retrain the model on serving data.
   • BApply log transformation to `call_duration`, use target encoding for `plan_type`, and rely on BigQuery ML to handle serving transformations.
   • CRemove outliers from `call_duration`, label encode `plan_type`, and manually transform serving data.
   • DApply robust scaling to `call_duration`, one-hot encode `plan_type`, and create a serving pipeline that applies the same transformations.
   Reveal Solution Discussion Select an answer first
4. 18application · medium

   A social media analytics company wants to generate embeddings for user posts to detect trending topics. They have a mix of short tweets and long blog posts. The embedding model they use has a maximum token limit of 512. What is the best approach to generate embeddings for this mixed-length content?

   • AUse a different model for short and long content.
   • BTruncate all content to 512 tokens before embedding.
   • CChunk long posts into segments of 512 tokens with overlap, and embed each chunk separately.
   • DGenerate a single embedding for the entire post, ignoring the token limit.
   Reveal Solution Discussion Select an answer first
5. 19application · medium

   A legal tech company is building a RAG system to answer questions from a large corpus of court rulings. They have already chunked the documents and generated embeddings. Now they need to index the embeddings in a vector store. Which additional step is essential to ensure high-quality retrieval?

   • AApply a hash function to the embeddings to reduce storage size.
   • BStore metadata such as case number, date, and jurisdiction alongside each embedding.
   • CStore only the original text and skip the vector index.
   • DUse a relational database to store the embeddings in a single column.
   Reveal Solution Discussion Select an answer first
6. 20expert · hard

   A music streaming company wants to build a recommendation system using embeddings of song audio features. They have millions of songs and plan to use a pre-trained model to generate embeddings. The embeddings will be stored in a vector database for similarity search. The data engineering team needs to decide on a storage and indexing strategy that balances query latency and cost. Which approach is most appropriate?

   • AUse a key-value store without any indexing.
   • BStore embeddings in a relational database and use brute-force similarity search.
   • CStore embeddings in a CSV file and load them into memory for each query.
   • DUse a dedicated vector database with approximate nearest neighbor (ANN) indexing to support low-latency queries at scale.
   Reveal Solution Discussion Select an answer first