Prueba RAG correctiva (CRAG)

Las respuestas y explicaciones aparecen a medida que realiza el cuestionario.

1. What core weakness of standard RAG does CRAG address?
   • The generator trusts whatever the retriever returns
   • Embeddings are too small
   • Models cannot read long documents
   • Vector databases are too slow
2. In the original CRAG paper, what kind of model served as the retrieval evaluator?
   • A large GPT-style generator
   • A fine-tuned T5-large model
   • A BM25 index
   • A human reviewer
3. What does CRAG do when all retrieved documents score low?
   • Answers from the documents anyway
   • Returns an error to the user
   • Discards them, rewrites the query and uses web search
   • Doubles the number of retrieved documents from the same index
4. What is the purpose of the decompose-then-recompose step?
   • To translate documents into another language
   • To compress embeddings
   • To rank web pages by popularity
   • To split documents into strips and keep only relevant pieces
5. When the evaluator is unsure about relevance, which action does CRAG take?
   • Ambiguous: combine refined internal knowledge with web results
   • Correct: use internal documents only
   • Incorrect: use web results only
   • Stop and ask the user to rephrase
6. Which statement about CRAG is a misconception according to the guide?
   • It adds cost and latency
   • It can wrap other RAG methods
   • CRAG verifies that the final answer is correct
   • It can use web search as a fallback
7. What is the trade-off of setting a very strict relevance threshold?
   • Weak documents slip through more often
   • Fallback triggers often, adding web calls and latency
   • The system stops using an evaluator
   • Answers become shorter
8. Why log which path (Correct, Ambiguous, Incorrect) each query took?
   • It is required by the CRAG license
   • It trains the embedding model automatically
   • It removes the need for thresholds
   • It reveals coverage gaps in the knowledge base worth fixing