GraphQLConf 2024

Discovering GraphQL fields and generating queries is a tedious task for developers. They spend a considerable amount of time finding the appropriate fields in large schemas. To solve this problem at Intuit where we have a super graph consisting of millions of lines we implemented a framework that makes use of Generative AI to help developers with attribute discovery and query generation. The benefits of our approach include being able to work with large schemas without the hassle of going through the whole schema and requiring less back and forth communication between consumers and schema owners, which results in a huge boost in developer productivity. We created chunks of the schema and ingested them into a vector store, we then do a retrieval, dynamically build a minimal schema and perform RAG where the LLM is provided with the minimal schema and the user query. The LLM responds with either a list of discovered attributes or GraphQL query. This framework aims to achieve lower latency and less hallucinations by reducing the size of the schema sent to the LLM, this also results in lower costs and higher accuracy.