1.5 Describe Retrieval Augmented Generation (RAG) and Role of Embeddings and Vector Databases AI-TECHNICAL-PRACTITIONER Practice Questions (Page 1)

Questions 1–5

1. 1expert · hard

   A team is building a RAG system for a niche legal domain. They use a general-purpose embedding model and find that retrieval quality is poor for domain-specific terms (e.g., 'force majeure', 'indemnification'). They have a small set of 5,000 labeled query-document pairs. Which approach is most likely to improve retrieval quality?

   • ASwitch to a larger general-purpose embedding model without any domain adaptation.
   • BIncrease the number of retrieved chunks from 5 to 20 to compensate for poor embeddings.
   • CUse a keyword-based search instead of embeddings for all queries.
   • DFine-tune the embedding model on the labeled pairs using contrastive learning, so domain-specific terms are mapped closer to their relevant documents.
   Reveal Solution Discussion Select an answer first
2. 2foundation · easy

   What is the primary role of a vector database in a RAG system?

   • ATo train the embedding model on the domain-specific data
   • BTo store the original text documents in a relational format
   • CTo store embeddings and efficiently retrieve vectors similar to a query vector
   • DTo generate the final response by running the language model
   Reveal Solution Discussion Select an answer first
4. 3foundation · easy

   What is a key difference between RAG and fine-tuning in terms of updating a model's knowledge?

   • ARAG can incorporate new information without retraining, while fine-tuning requires updating the model's parameters
   • BRAG updates the model's weights, while fine-tuning does not
   • CRAG requires more training data than fine-tuning
   • DRAG is only used for text generation, while fine-tuning is only used for classification
   Reveal Solution Discussion Select an answer first
5. 4application · medium

   A team is evaluating two embedding models for a RAG system. Model X produces 384-dimensional vectors; Model Y produces 1536-dimensional vectors. Both models are otherwise similar in quality. The team's vector database stores 10 million vectors and query latency is a concern. What is the most likely trade-off they will observe?

   • AModel Y will be faster because higher dimensions allow the vector database to use a simpler index.
   • BModel X will always produce more accurate retrieval results because lower dimensions reduce overfitting.
   • CBoth models will produce identical retrieval results because embeddings are normalized to unit length.
   • DModel Y will use more storage and may have slower queries, but may capture finer semantic nuances than Model X.
   Reveal Solution Discussion Select an answer first
6. 5application · medium

   A team is building a RAG system and is deciding how to split a 100-page technical manual into chunks. They observe that retrieval quality is poor: answers are incomplete because the relevant information is split across two adjacent chunks. Which chunking strategy is most likely to fix this?

   • AUse smaller chunks so each chunk contains less information and is more focused.
   • BUse larger chunks so each chunk contains more information, even if it exceeds the embedding model's token limit.
   • CUse overlapping chunks so that the boundary between chunks includes some context from the neighboring section.
   • DRemove chunking entirely and embed the entire manual as a single vector.
   Reveal Solution Discussion Select an answer first