Week 12: Retrieval Augmented Generation — Knowledge Check

Question 1

A) Knowledge cutoffs that prevent access to information after the training date
B) Inability to access private or proprietary data at inference time
C) Tendency to hallucinate plausible-sounding but incorrect facts
D) Inability to process multiple languages without separate fine-tuning per language

Question 2

A) Generate draft response → Embed query → Retrieve documents → Refine response
B) Retrieve documents → Embed query → Augment prompt → Generate response
C) Embed query → Search vector database → Retrieve relevant documents → Augment prompt with retrieved context → Generate response
D) Embed query → Generate response → Retrieve documents → Post-process with context

Question 3

A) It requires storing two copies of every vector for comparison purposes
B) It requires O(n * d) operations where n is the number of vectors and d is the dimensionality, becoming too slow at scale
C) It can only compute cosine similarity and fails with Euclidean distance metrics
D) It requires in-memory storage of the entire dataset, which exceeds GPU memory limits

Question 4

A) It hashes similar vectors to the same buckets and only compares vectors within matching buckets
B) It compresses vectors by quantizing subvectors with learned codebooks, reducing memory and compute
C) It builds a multi-layer graph where search starts at the top layer with long-range connections and progressively narrows through layers with finer granularity
D) It pre-sorts all vectors by their first principal component and performs binary search on the sorted index

Question 5

A) To increase the total number of chunks so that more candidates are retrieved per query
B) To ensure that information near chunk boundaries is not lost between adjacent chunks
C) To prevent the embedding model from receiving identical input sequences
D) To make adjacent chunk embeddings more similar so they are always retrieved together

Question 6

A) Cross-encoders are faster than bi-encoders and can rank larger candidate sets more efficiently
B) Cross-encoders process the query and document jointly, capturing fine-grained interactions that separate embeddings miss
C) Cross-encoders reduce embedding dimensionality, lowering storage costs in the vector database
D) Cross-encoders generate query-specific document embeddings that replace the static pre-computed ones

Question 7

A) Queries are typically longer than documents and must be compressed into answer-length representations before comparison
B) A hypothetical answer to the query is semantically closer to relevant documents in the embedding space than the query itself
C) Generating a hypothetical answer first allows the system to skip retrieval entirely for simple questions
D) It forces the model to consult its parametric knowledge before the external retrieval step

Question 8

A) Recall@k measures how fast the retriever returns results; Precision@k measures how accurate the results are
B) Recall@k measures the proportion of all relevant documents that appear in the top-k; Precision@k measures the proportion of top-k results that are relevant
C) Recall@k and Precision@k are identical metrics with different names used by different research communities
D) Recall@k measures the faithfulness of the generated response; Precision@k measures the relevance of retrieved documents

Question 9

A) Dense embeddings cannot be stored in vector databases, so a sparse fallback is required
B) Semantic search can miss results for rare or specialized terms where exact keyword matching is more reliable, so combining both improves robustness
C) BM25 is faster than vector search and is used as a caching layer to reduce embedding model calls
D) Dense embeddings work only for questions, while BM25 is needed for document retrieval

Question 10

A) Whether the response uses the same vocabulary and phrasing as the retrieved documents
B) Whether every claim in the generated response is supported by the retrieved context, rather than drawn from the model's parametric memory
C) The semantic similarity between the user query and the generated response
D) The proportion of retrieved documents that were actually used in generating the response

Week 12: Retrieval Augmented Generation (RAG)