For AI and database optimization problems, existing technologies need to change the bottom layer of database, which affects the application of research results and lacks scalability. A learning query optimization method for non-embedded database was proposed. In the cardinality estimation stage, the multi-model method is used to establish a neural network for specific sub queries and train different sub models independently, which solves the problem of too many training sets and poor scalability. In the join optimization stage, cost-based reinforcement learning is applied to improve the query optimization performance. For each query, the optimization processes from cardinality estimation to connection sorting are executed outside the database. The query is rewritten according to the obtained optimization strategy, and the rewriting results are returned to the database. The query is executed according to the specified plan by setting parameters. Experimental verification was carried out on the data set containing eight tables. Compared with the query not optimized, the optimization method of non-embedded database has good optimization effect.