Dialogue system is one of the research area coming into picture because of advancement in natural language processing and deep learning methods. Dialogue systems are designed for communication between humans and machine. When humans communicate with each other they use their own intelligence to carry conversation but this intelligence is missing in machines. Researchers have attempted to accommodate external knowledge with machines to generate knowledge-enhanced responses. Knowledge graph is one of the structured ways of providing an abstraction of the real world knowledge to the machine, and machine in turn can use this knowledge to improve the quality of response generated by dialogue systems. Generating knowledge grounded response is a challenging task. Recently most of the architectures are end-to-end dialogue system, in contrast to them this paper proposes three step architecture which extracts entity from input using inside outside beginning 2 tagging and Bidirectional Encoder Representations from Transformers, secondly entity related sub-graph is extracted using laplacian matrix method then knowledge grounded response are generated using extracted subgraph and Gated recurrent unit encoder-decoder model. This architecture has independent fact retrieval system which is detached from two tunable NER model and response generation model which makes the model training easy as compared to end-to-end trainable system and also improves the overall performance of the system. Proposed model is tested on standard benchmark dataset In-car and shows performance comparable with existing models.