One assumption in an IR system is that all documents in a collection are indexed. Thus it is possible to find all relevant documents, and know that all documents have been found, given the proper query. In the field of Artificial Intelligence, this is known as the Closed World Assumption. Relevance Feedback techniques are designed with the assumption that all documents are indexed, and that their purpose is to find relevant documents that have been indexed, but have not yet been returned to the user.
In many cases, in particular the World Wide Web, the document index is often incomplete. Thus, it may be possible for an IR system to return all relevant documents that have been indexed, but not all relevant documents in the collection. In order for Relevance Feedback to work, it will likely need to engage in search through unindexed documents. This assumption alters how Relevance Feedback operates more than any I've stated. By adding a requirement that a RF technique search through unindexed documents, there are now questions as to which documents to search, in what order, and at what cost. There has been some work on the Web that implements a system to search for relevant documents on the web given some relevance information [26], however this needs to be evaluated in the context of Relevance Feedback.