In open-domain Question Answering, the Retriever-Reader pipeline works as follows. Which statements are correct?

Question

Accepted Answer

Dense retrieval methods encode questions and documents as vectors and find the closest match

Accepted Answer

The retriever finds relevant documents from a large collection given the question

Accepted Answer

The reader then extracts the answer span from the retrieved documents

Answer

The retriever reads every document fully with the QA model before ranking them

Crowdly