From a9896cd0727215c9f140fce8f60b61ca067b961b Mon Sep 17 00:00:00 2001
From: spped2000 <spped2000@gmail.com>
Date: Wed, 4 Nov 2026 20:43:08 +0700
Subject: [PATCH] fix: keep non-ASCII in prompt schemas (#1130)

StructuredResponseRecipe.schema and PydanticResponseRecipe.schema
serialize the JSON schema with json.dumps' default ensure_ascii=True.
The schema is embedded in the user prompt, so non-ASCII descriptions,
enum values and LLM-judge rubrics (Thai, CJK, accented Latin, ...)
reach the model as \uXXXX escapes instead of readable text.

Serialize with ensure_ascii=False, as serialize_output in the same
file already does. ASCII-only schemas are unchanged.
generate_response_example gets the same change for consistency. Tests
cover Thai descriptions and enum values and check that an ASCII-only
schema serializes exactly as before.

Fixes #1130

Signed-off-by: spped2000 <spped2000@gmail.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---
 .../engine/models/recipes/response_recipes.py |  6 +--
 .../models/recipes/test_response_recipes.py   | 41 +++++++++++++++++++
 2 files changed, 44 insertions(+), 3 deletions(-)

diff --git a/packages/data-designer-engine/src/data_designer/engine/models/recipes/response_recipes.py b/packages/data-designer-engine/src/data_designer/engine/models/recipes/response_recipes.py
index e32f212cf..675303f6b 100644
--- a/packages/data-designer-engine/src/data_designer/engine/models/recipes/response_recipes.py
+++ b/packages/data-designer-engine/src/data_designer/engine/models/recipes/response_recipes.py
@@ -106,11 +106,11 @@ def example_template(self) -> str:
         return "```json\n{example}\n```"
 
     def generate_response_example(self, example: dict) -> str:
-        return self.example_template.format(example=json.dumps(example))
+        return self.example_template.format(example=json.dumps(example, ensure_ascii=False))
 
     @property
     def schema(self) -> str:
-        return json.dumps(self.json_schema)
+        return json.dumps(self.json_schema, ensure_ascii=False)
 
     def serialize_output(self, output: dict) -> str:
         return json.dumps(output, ensure_ascii=False)
@@ -182,7 +182,7 @@ def __init__(self, data_type: type[BaseModel], **kwargs):
 
     @property
     def schema(self) -> str:
-        return json.dumps(self.data_type.model_json_schema())
+        return json.dumps(self.data_type.model_json_schema(), ensure_ascii=False)
 
     @property
     def task_instructions(self) -> str:
diff --git a/packages/data-designer-engine/tests/engine/models/recipes/test_response_recipes.py b/packages/data-designer-engine/tests/engine/models/recipes/test_response_recipes.py
index de868fd54..6482af326 100644
--- a/packages/data-designer-engine/tests/engine/models/recipes/test_response_recipes.py
+++ b/packages/data-designer-engine/tests/engine/models/recipes/test_response_recipes.py
@@ -322,6 +322,47 @@ def test_structured_response_task_instructions():
     assert "Baz!!" in recipe.task_instructions
 
 
+class ThaiFoo(BaseModel):
+    province: str = Field(..., description="ชื่อจังหวัด")
+    region: str = Field(..., description="ภาค", json_schema_extra={"enum": ["ภาคเหนือ", "ภาคใต้"]})
+
+
+class AsciiFoo(BaseModel):
+    province: str = Field(..., description="Province name")
+    region: str = Field(..., description="Region", json_schema_extra={"enum": ["North", "South"]})
+
+
+@pytest.mark.parametrize(
+    "recipe",
+    [PydanticResponseRecipe(ThaiFoo), StructuredResponseRecipe(ThaiFoo.model_json_schema())],
+    ids=["pydantic", "structured"],
+)
+def test_json_recipe_task_instructions_keep_non_ascii_text(
+    recipe: PydanticResponseRecipe | StructuredResponseRecipe,
+) -> None:
+    assert "ชื่อจังหวัด" in recipe.task_instructions
+    assert "ภาคเหนือ" in recipe.task_instructions
+    assert "\\u0e" not in recipe.task_instructions
+    assert json.loads(recipe.schema) == ThaiFoo.model_json_schema()
+
+
+@pytest.mark.parametrize(
+    "recipe",
+    [PydanticResponseRecipe(AsciiFoo), StructuredResponseRecipe(AsciiFoo.model_json_schema())],
+    ids=["pydantic", "structured"],
+)
+def test_json_recipe_schema_unchanged_for_ascii(recipe: PydanticResponseRecipe | StructuredResponseRecipe) -> None:
+    assert recipe.schema == json.dumps(AsciiFoo.model_json_schema())
+
+
+def test_structured_response_example_keeps_non_ascii_text() -> None:
+    recipe = StructuredResponseRecipe(ThaiFoo.model_json_schema())
+    example = {"province": "เชียงใหม่", "region": "ภาคเหนือ"}
+
+    assert "เชียงใหม่" in recipe.generate_response_example(example)
+    assert recipe.parse(recipe.generate_response_example(example)) == example
+
+
 def test_pydantic_response_patched_case_single_lb():
     recipe = PydanticResponseRecipe(Foo)
     example = Foo(bar=Bar(baz=42))