In recent years, there has been a lot of interest in the field of keyword querying relational databases. A variety of systems such as DBXplorer [ACD02], Discover [HP02] and ObjectRank [BHP04] have been proposed. Another such system is BANKS, which enables data and schema browsing together with keyword-based search for relational databases. It models tuples as nodes in a graph, connected by links induced by foreign key and other relationships. The size of the database graph that BANKS uses is proportional to the sum of the number of nodes and edges in the graph. Systems such as SPIN, which search on Personal Information Networks and use BANKS as the backend, maintain a lot of information about the users' data. Since these systems run on the user workstation which have other demands of memory, such a heavy use of memory is unreasonable and if possible, should be avoided. In order to alleviate this problem, we introduce EMBANKS (acronym for External Memory BANKS), a framework for an optimized disk-based BANKS system. The complexity of this framework poses many questions, some of which we try to answer in this thesis. We demonstrate that the cluster representation proposed in EMBANKS enables in-memory processing of very large database graphs. We also present detailed experiments that show that EMBANKS can significantly reduce database load time and query execution times when compared to the original BANKS algorithms.