donkit-llm 0.1.5__tar.gz → 0.1.6__tar.gz

{donkit_llm-0.1.5 → donkit_llm-0.1.6}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.3
 Name: donkit-llm
-Version: 0.1.5
+Version: 0.1.6
 Summary: Unified LLM model implementations for Donkit (OpenAI, Azure OpenAI, Claude, Vertex AI, Ollama)
 License: MIT
 Author: Donkit AI

{donkit_llm-0.1.5 → donkit_llm-0.1.6}/pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "donkit-llm"
-version = "0.1.5"
+version = "0.1.6"
 description = "Unified LLM model implementations for Donkit (OpenAI, Azure OpenAI, Claude, Vertex AI, Ollama)"
 authors = ["Donkit AI <opensource@donkit.ai>"]
 license = "MIT"

{donkit_llm-0.1.5 → donkit_llm-0.1.6}/src/donkit/llm/openai_model.py

@@ -43,6 +43,43 @@ class OpenAIModel(LLMModelAbstract):
         self._init_client(api_key, base_url, organization)
         self._capabilities = self._determine_capabilities()
 
+    def _clean_schema_for_openai(self, schema: dict, is_gpt5: bool = False) -> dict:
+        """Clean JSON Schema for OpenAI strict mode.
+        
+        Args:
+            schema: JSON Schema to clean
+            is_gpt5: Currently unused - both GPT-4o and GPT-5 support title/description
+        
+        OpenAI Structured Outputs supports:
+            - title and description (useful metadata for the model)
+            - Automatically adds additionalProperties: false for all objects
+            - Recursively processes nested objects and arrays
+        
+        Note: The main requirement is additionalProperties: false for all objects,
+        which is automatically added by this method.
+        """
+        cleaned = {}
+        
+        # Copy all fields, recursively processing nested structures
+        for key, value in schema.items():
+            if key == "properties" and isinstance(value, dict):
+                # Recursively clean properties
+                cleaned["properties"] = {
+                    k: self._clean_schema_for_openai(v, is_gpt5) if isinstance(v, dict) else v
+                    for k, v in value.items()
+                }
+            elif key == "items" and isinstance(value, dict):
+                # Recursively clean array items
+                cleaned["items"] = self._clean_schema_for_openai(value, is_gpt5)
+            else:
+                cleaned[key] = value
+        
+        # Ensure additionalProperties is false for objects (required by OpenAI)
+        if cleaned.get("type") == "object" and "additionalProperties" not in cleaned:
+            cleaned["additionalProperties"] = False
+            
+        return cleaned
+
     def _get_base_model_name(self) -> str:
         """Get base model name for capability/parameter detection.
 
@@ -267,13 +304,18 @@ class OpenAIModel(LLMModelAbstract):
             # If response_format is a JSON Schema dict with "type": "object", wrap it
             if isinstance(request.response_format, dict):
                 if request.response_format.get("type") == "object":
-                    # This is a JSON Schema - wrap it in json_schema format
+                    # This is a JSON Schema - clean and wrap it in json_schema format
+                    # GPT-5 and reasoning models need stricter schema cleaning
+                    is_gpt5 = self._is_reasoning_model()
+                    cleaned_schema = self._clean_schema_for_openai(
+                        request.response_format, is_gpt5=is_gpt5
+                    )
                     kwargs["response_format"] = {
                         "type": "json_schema",
                         "json_schema": {
                             "name": "response",
                             "strict": True,
-                            "schema": request.response_format,
+                            "schema": cleaned_schema,
                         },
                     }
                 else: