Hi friends! For defining function calls input, I wonder if anyone had experience with GraphQL schema instead of JSON schema for input parameters? I assume ultimately it comes down how good the LLM is at generating good GraphQL to get the job done, so I'm curious if anyone had tried and if so, which models did you use? The GraphQL schema option is especially appealing to us because: 1) seems more concise than JSON schema; 2) allows to skip unnecessary fields in the response which really cuts down the token usage.
I've wondered the same thing. Feels LLMs are so biased to work via JSON schema I've kinda given up on GraphQL for LLMs - notably because I've started advertising tools via MCP which really feels more like tools over RPC. But letting an LLM synthesize graphql as a single tool makes so much sense. LMK if you find a good solution! Because that would be so powerful.
Thanks for the input Mikyo! Like you said, it'll be so powerful as graphql does introspection and can self discover the available tools. We are defining tool calls for our internal services which are all gRPC services. We'll start with JSON schema and then experiment with GraphQL schema as we go!
