Speculative RAG and Retrieval-Augmented Drafting Quiz

Answers and explanations appear as you take the quiz.

1. In Speculative RAG, what role does the smaller model play?
   • It drafts multiple candidate answers from retrieved document subsets
   • It verifies and ranks final answers
   • It stores the vector database
   • It only handles tokenization
2. Why are retrieved documents clustered before drafting?
   • To compress them into a single token
   • To form diverse, non-redundant evidence subsets for different drafts
   • To delete duplicate models
   • To increase the model's vocabulary
3. What does the larger 'verifier' model primarily do?
   • Generate the retrieved documents
   • Train the embedding model
   • Score the candidate drafts and select the best one
   • Cluster the documents
4. How does Speculative RAG reduce latency compared to standard RAG?
   • By skipping retrieval altogether
   • By using a smaller context window with no documents
   • By caching the user's question
   • By having a small model do parallel drafting so the big model only checks short drafts
5. Each draft produced by the drafter typically includes:
   • A rationale explaining the answer
   • A new set of model weights
   • The full vector index
   • A list of all users
6. Speculative RAG's draft-then-verify structure is conceptually similar to which decoding technique?
   • Beam search pruning
   • Speculative decoding
   • Greedy sampling
   • Top-k filtering
7. What information can the Speculative RAG verifier use when ranking candidate drafts?
   • Only the length of each document
   • The GPU type and cluster name
   • Draft consistency, rationale likelihood and a self-reflection signal
   • A random choice among the drafts
8. Why sample documents from different clusters for separate drafts?
   • To ensure every draft sees identical passages
   • To remove all retrieved evidence
   • To make each drafter use a different tokenizer
   • To give the drafts diverse, less-redundant evidence subsets