Design Retrieval Systems Using Advanced Chunking Strategies GENERATIVE-AI-ENGINEER-ASSOCIATE Practice Questions (Page 4)

Questions 16–20

1. 16application · medium

   A data science team has implemented two chunking strategies for a RAG system: fixed-size chunking with 300 tokens and semantic chunking. They want to objectively decide which one to deploy. They have a labeled test set of questions with relevant document passages. Which evaluation approach should they use?

   • AAsk a few users to manually review the chunks and give subjective feedback.
   • BCompare the total number of chunks produced by each strategy.
   • CMeasure retrieval recall and precision on the labeled test set for both strategies and compare.
   • DMeasure the time taken to generate chunks for the entire corpus.
   Reveal Solution Discussion Select an answer first
2. 17application · medium

   A university library is digitizing textbooks and wants to build a RAG system for students. The textbooks have clear chapters, sections, and subsections. Students often ask questions that reference a specific section, like 'What does section 3.2 say about photosynthesis?'. The team wants chunks that align with the book's structure so that retrieval results are easy to cite. Which chunking strategy should they use?

   • ARecursive character text splitting with paragraph separators.
   • BSemantic chunking to group related content across sections.
   • CFixed-size chunking with a small chunk size to capture details.
   • DDocument-based chunking that splits on headings and subheadings.
   Reveal Solution Discussion Select an answer first
4. 18expert · hard

   A team is building a RAG system over a corpus of academic papers. They have a limited budget for embedding computation. They are considering semantic chunking, which produces high-quality chunks but is computationally expensive. They also have a fixed-size chunking baseline. They want to decide whether the improved retrieval quality justifies the cost. What should they do?

   • ARun a small-scale evaluation comparing retrieval metrics like recall@k and precision@k on a sample of queries.
   • BAssume semantic chunking is always better and deploy it.
   • CUse semantic chunking only for a subset of documents and fixed-size for the rest.
   • DUse fixed-size chunking because it is cheaper and good enough.
   Reveal Solution Discussion Select an answer first
5. 19application · medium

   A media company is building a RAG system over long-form articles. They notice that when a chunk is retrieved, the model sometimes lacks the context of what came before, leading to answers that are out of context. They want to retain surrounding context without making chunks too large. Which technique should they use?

   • AUse document-based chunking by paragraphs.
   • BUse a sliding window approach where each chunk includes a portion of the previous chunk's text.
   • CUse semantic chunking to group all related sentences into one large chunk.
   • DUse fixed-size chunking with no overlap to keep chunks small.
   Reveal Solution Discussion Select an answer first
6. 20application · medium

   A startup is building a RAG system over a large corpus of user-generated content, such as forum posts. The posts vary greatly in length and quality. They need a simple, fast, and predictable chunking method that ensures each chunk is roughly the same size for consistent embedding and retrieval. Which method should they use?

   • ARecursive character text splitting.
   • BDocument-based chunking by post.
   • CFixed-size chunking with a configurable chunk size and overlap.
   • DSemantic chunking.
   Reveal Solution Discussion Select an answer first