Online social networks are rapidly becoming popular for user to share, organize and locate interesting content. However, due to the privacy control, users can only access a fraction of the whole networks, e.g., those direct of two-hop friends. Due to a large amount of possible relevant data of two-hop friends, users commonly browse the profiles and the homepages, which dominate the 92% user behaviors, and are inefficient in obtaining desired information for a user. In this paper, we propose an efficient keyword search model to help users achieve desired information. We propose a novel summary index with a ranking model by extending existing Bloom filter techniques, and achieve efficient full-text search over large scale OSNs to reduce inter-server communication cost and provide much shorter query latency. Furthermore, we conduct comprehensive simulations using traces from real world systems to evaluate our design. Results show that our scheme reduces the network traffic by 94% and reduces the query latency by 82% with high search accuracy.