Schema manifesto

[iboB]

Every author of an inference provider is free to choose their own API. The selected API design can follow existing conventions or choose their own. A formal schema would greatly help in establishing and following conventions.

Moreover every API user would prefer to use strongly-typed code which follows their language's conventions rather than composing and reading frames from raw bytes. Even with our current use of the Dict helper, we still prefer using the C++ schema definition when we write examples, demos, and tests. A schema can be used to generate code.

Documenting an API can be hard and cumbersome. There is a risk in divergence between the implementation and documentation. A schema can be used to generate the docs, and if it's also used to generate the code, the risk of divergence becomes zero.

We need to invest in schemas: codegen, documentation generation, tooling