Idea: Jsonschema parser for functions?

[eloycoto]

When working with functions in PDL, there is a small disadvantage: you need to declare the function twice to use it. Once for PDL consumption, and another time to attach it to the tools section in the LLM call.

I was thinking of adding a new parser that takes advantage of:

prompt-declaration-language/src/pdl/pdl_schema_utils.py

Lines 32 to 76 in ad048b8

	def pdltype_to_jsonschema( # pylint: disable=too-many-return-statements
	pdl_type: str | dict[str, Any] | list, additional_properties: bool
	) -> dict[str, Any]:
	match pdl_type:
	case {"enum": choices}:
	return {"enum": choices}
	case None:
	return {"type": "null"}
	case "bool" | "str" | "float" | "int" | "list" | "obj":
	return {"type": _PDLTYPE_TO_JSONSCHEMA_NAME[pdl_type]}
	case {"str": dict() as details}:
	return {"type": "string", **details}
	case {"float": dict() as details}:
	return {"type": "number", **details}
	case {"int": dict() as details}:
	return {"type": "integer", **details}
	case {"list": str() as type_name}:
	return {
	"type": "array",
	"items": pdltype_to_jsonschema(type_name, additional_properties),
	}
	case {"list": dict() as details}:
	ikws = ["enum", *_PDLTYPE_TO_JSONSCHEMA_NAME.keys()]
	items_details = {k: v for k, v in details.items() if k in ikws}
	if len(items_details) != 1:
	raise ValueError(f"invalid PDL type {pdl_type}")
	other_details = {k: v for k, v in details.items() if k not in ikws}
	return {
	"type": "array",
	"items": pdltype_to_jsonschema(items_details, additional_properties),
	**other_details,
	}
	case list() as type_list:
	if len(type_list) != 1:
	raise ValueError(f"invalid PDL type {pdl_type}")
	return {
	"type": "array",
	"items": pdltype_to_jsonschema(type_list[0], additional_properties),
	}
	case {"obj": dict() as pdl_props}:
	return get_json_schema_object(pdl_props, additional_properties)
	case dict() as pdl_props:
	return get_json_schema_object(pdl_props, additional_properties)
	raise ValueError(f"invalid PDL type {pdl_type}")

This way, the PDL script could look like this:

description: test
defs:
  get_current_stock:
    function:
      product_name: str
    return:
      "FN::get_current_stock:: '${product_name}'\n"
text:
  - def: stock_fn_jsonschema
    parser: jsonschema
    return:
      "${get_current_stock}"

Here, you define the jsonschema, and tools can be defined in an array using a for loop with all available PDL functions. The PDL script will be easier to read and feel more natural.

A small change to the PDL schema is needed, such as adding descriptions for arguments and functions, so it might be more complicated than it initially seems due to changes in the schema.

[vazirim]

@eloycoto This would be great! When you make the change to the AST (src/pdl/pdl_ast.pdl), you can run the following command to regenerate the PDL schema:

pdl --schema > src/pdl/pdl-schema.json

Also follow instructions that are here to update the Live Viewer:
https://github.com/IBM/prompt-declaration-language/tree/main/pdl-live

Thanks!