[Feature] Allow oneOf in JSON schemas (with limited support)

guidance/library/_json.py

@@ -11,6 +11,7 @@
     Type,
     TYPE_CHECKING,
 )
+import warnings
 
 try:
     import jsonschema
@@ -39,6 +40,7 @@ def _to_compact_json(target: Any) -> str:
 class Keyword(str, Enum):
     ANYOF = "anyOf"
     ALLOF = "allOf"
+    ONEOF = "oneOf"
     REF = "$ref"
     CONST = "const"
     ENUM = "enum"
@@ -55,6 +57,7 @@ class Keyword(str, Enum):
 IGNORED_KEYS = {
     "$schema",
     "$id",
+    "id",
     "$comment",
     "title",
     "description",
@@ -63,6 +66,14 @@ class Keyword(str, Enum):
     "required",  # TODO: implement and remove from ignored list
 }
 
+# discriminator is part of OpenAPI 3.1, not JSON Schema itself
+# https://json-schema.org/blog/posts/validating-openapi-and-json-schema
+# TODO: While ignoring this key shouldn't lead to invalid outputs, forcing
+# the model to choose the value of the marked field before other fields
+# are generated (statefully or statelessly) would reduce grammar ambiguity
+# and possibly improve quality.
+IGNORED_KEYS.add("discriminator")
+
 TYPE_SPECIFIC_KEYS = {
     "array": {"items", "prefixItems", "minItems", "maxItems"},
     "object": {"properties", "additionalProperties"},
@@ -339,6 +350,13 @@ def _gen_json(
             raise ValueError("Only support allOf with exactly one item")
         return lm + _gen_json(allof_list[0], definitions)
 
+    if Keyword.ONEOF in json_schema:
+        oneof_list = json_schema[Keyword.ONEOF]
+        if len(oneof_list) == 1:
+            return lm + _gen_json(oneof_list[0], definitions)
+        warnings.warn("oneOf not fully supported, falling back to anyOf. This may cause validation errors in some cases.")
+        return lm + _process_anyOf(anyof_list=oneof_list, definitions=definitions)
+
     if Keyword.REF in json_schema:
         return lm + _get_definition(reference=json_schema[Keyword.REF], definitions=definitions)
 

tests/unit/library/test_json.py

@@ -8,7 +8,7 @@
 from guidance import json as gen_json
 from guidance import models
 
-from guidance.library._json import _to_compact_json, WHITESPACE
+from guidance.library._json import _to_compact_json, WHITESPACE, IGNORED_KEYS
 
 from ...utils import check_match_failure as _check_match_failure
 from ...utils import check_run_with_temperature
@@ -1319,6 +1319,37 @@ def test_allOf_bad_schema(self):
             lm += gen_json(name=CAPTURE_KEY, schema=schema_obj)
         assert ve.value.args[0] == "Only support allOf with exactly one item"
 
+class TestOneOf:
+    @pytest.mark.parametrize("target_obj", [123, 42])
+    def test_oneOf_simple(self, target_obj):
+        schema = """{
+            "oneOf" : [{ "type": "integer" }]
+        }
+        """
+        # First sanity check what we're setting up
+        schema_obj = json.loads(schema)
+        validate(instance=target_obj, schema=schema_obj)
+
+        # The actual check
+        generate_and_check(target_obj, schema_obj)
+
+
+    @pytest.mark.parametrize("target_obj", [123, True])
+    def test_oneOf_compound(self, target_obj):
+        schema = """{
+        "oneOf" : [{ "type": "integer" }, { "type": "boolean" }]
+        }
+        """
+        # First sanity check what we're setting up
+        schema_obj = json.loads(schema)
+        validate(instance=target_obj, schema=schema_obj)
+
+        # The actual check; we expect a warning here because oneOf is not fully supported
+        with pytest.warns() as record:
+            generate_and_check(target_obj, schema_obj)
+        assert len(record) == 1
+        assert record[0].message.args[0].startswith("oneOf not fully supported")
+
 
 class TestEnum:
     simple_schema = """{
@@ -1950,3 +1981,13 @@ def test_no_additionalProperties(self, compact):
             maybe_whitespace=True,
             compact=compact,
         )
+
+def test_ignored_keys_allowed_as_properties():
+    schema_obj = {
+        "type": "object",
+        "properties": {
+            key: {"type": "string"} for key in IGNORED_KEYS
+        }
+    }
+    target_obj = {key: "value" for key in IGNORED_KEYS}
+    generate_and_check(target_obj, schema_obj)

tests/unit/library/test_pydantic.py

@@ -274,3 +274,46 @@ def test_bad_generic(
             maybe_whitespace=maybe_whitespace,
             compact=compact,
         )
+
+class TestDiscriminatedUnion:
+    """
+    https://docs.pydantic.dev/latest/concepts/unions/#discriminated-unions-with-str-discriminators
+    """
+
+    class Cat(pydantic.BaseModel):
+        pet_type: Literal['cat']
+        meows: int
+
+
+    class Dog(pydantic.BaseModel):
+        pet_type: Literal['dog']
+        barks: float
+
+
+    class Lizard(pydantic.BaseModel):
+        pet_type: Literal['reptile', 'lizard']
+        scales: bool
+
+
+    class Model(pydantic.BaseModel):
+        pet: Union[
+            'TestDiscriminatedUnion.Cat',
+            'TestDiscriminatedUnion.Dog',
+            'TestDiscriminatedUnion.Lizard',
+        ] = pydantic.Field(..., discriminator='pet_type')
+        n: int
+
+    def test_good(self):
+        obj = {"pet": {"pet_type": "dog", "barks": 3.14}, "n": 42}
+        generate_and_check(obj, self.Model)
+
+    def test_bad(self):
+        check_match_failure(
+            bad_obj={"pet": {"pet_type": "dog"}, "n": 42},
+            good_bytes=b'{"pet":{"pet_type":"dog"',
+            failure_byte=b"}",
+            allowed_bytes={b","}, # expect a comma to continue the object with "barks"
+            pydantic_model=self.Model,
+            maybe_whitespace=False,
+            compact=True
+        )