GraphQL wasn't made for AI. But it might be one of the best ways to talk to it (chillicream.com)

A new approach, termed Semantic Introspection, is being proposed for GraphQL to improve interactions with large language models (LLMs), addressing challenges that traditional APIs like OpenAPI and MCP face in effectively serving AI. As the software landscape transitions from human users to AI agents, there's a pressing need for APIs to be discoverable, scalable, and precise to minimize input data costs and enhance response efficiency. Classic performance metrics have become less relevant, as the emphasis shifts to reducing the number of API calls and the amount of data sent in each call. Semantic Introspection introduces an innovative querying capability, allowing agents to search and retrieve relevant schema members based on user queries without overwhelming the LLM with excessive data. This advancement ensures that LLMs can efficiently discover features and utilize APIs without running into context pollution or cost issues associated with larger schemas. Early experiments show that this approach significantly reduces the number of tokens and associated costs compared to other solutions, making it a promising direction for the AI/ML community seeking more effective API interactions. The latest Hot Chocolate preview supports this feature, suggesting a shift toward enhanced integration of GraphQL with LLMs.