Cache-Augmented Generation (CAG) Quiz

Answers and explanations appear as you take the quiz.

1. What does CAG use instead of per-query retrieval?
   • A keyword search engine
   • Fine-tuning the model on the documents
   • A new vector index per question
   • Preloading the whole knowledge base into the KV cache once and reusing it
2. What is the KV cache in a transformer?
   • A database of user accounts
   • The stored attention keys and values computed for tokens the model has already read
   • A list of document URLs
   • The model's trained weights
3. Why does a CAG system truncate the cache after each answer?
   • To save the answer for training
   • To delete the documents permanently
   • To restore the documents-only state so the next query does not see the previous question and answer
   • To switch to a different model
4. What is CAG's main limitation?
   • The entire corpus must fit in the model's context window
   • It requires a vector database
   • It cannot answer questions quickly
   • It changes the model's weights
5. Why does CAG answer faster than pasting the documents into every request?
   • It skips generating output tokens
   • The model only needs to compute the new question tokens on top of the cached prefix
   • It uses a smaller model
   • It retrieves fewer documents
6. Which knowledge base is the best fit for CAG?
   • A news archive updated every hour
   • Millions of customer records with per-user permissions
   • An entire web crawl
   • A small, stable product manual that rarely changes
7. Which statement about CAG is a misconception?
   • CAG removes retrieval errors because all documents are present
   • Changing documents requires rebuilding the cache
   • CAG fine-tunes the model's weights on the documents
   • Hosted prompt caching is related to the idea
8. Which techniques reduce the memory needed for a large KV cache?
   • Grouped-query attention and cache quantization
   • Adding more attention heads
   • Using longer prompts
   • Raising the sampling temperature