In the era of big data, there is a lot of information, but there is very little data that is really valuable to you. It is crucial to retrieve the data you really need from the redundant data. The traditional search system is only able to display the data that has appeared, not able to integrate, calculate and other processing of the data. This paper uses NL2SQL technology to implement the search system. First, we construct the heterogeneous graph structure to preprocess the data, then encode the heterogeneous graph into a high-dimensional hidden layer vector by the self-attention mechanism of relative position, and finally decode the high-dimensional hidden layer vector into SQL statements by the tree decoder. This can better meet the real needs of users, who can interact with the database more freely and fully exploit the value of the data.