Add support for json_mode unsupported models (#1239)

* Add support for json_mode unsupported models

* w -> __extract_json_from_text

* Update to version 0.0.95

Author: whoisarpit

patchwork/steps/SimplifiedLLM/SimplifiedLLM.py

@@ -16,6 +16,12 @@
 
 
 class SimplifiedLLM(Step):
+    # Models that don't support native JSON mode
+    JSON_MODE_UNSUPPORTED_MODELS = {
+        "gemini-2.0-flash-thinking-exp",
+        # Add other models here as needed
+    }
+
     def __init__(self, inputs):
         super().__init__(inputs)
         missing_keys = SimplifiedLLMInputs.__required_keys__.difference(set(inputs.keys()))
@@ -28,6 +34,7 @@ def __init__(self, inputs):
         self.is_json_mode = inputs.get("json", False)
         self.json_example = inputs.get("json_example")
         self.inputs = inputs
+        self.is_json_mode_unsupported = inputs.get("model") in self.JSON_MODE_UNSUPPORTED_MODELS
 
     def __record_status_or_raise(self, retry_data: RetryData, step: Step):
         if retry_data.retry_count == retry_data.retry_limit or step.status != StepStatus.FAILED:
@@ -49,15 +56,31 @@ def __json_loads(json_str: str) -> dict:
             logger.debug(f"Json to decode: \n{json_str}\nError: \n{e}")
             raise e
 
+    @staticmethod
+    def __extract_json_from_text(text: str) -> str:
+        try:
+            start = text.find("{")
+            end = text.rfind("}")
+            if start != -1 and end != -1:
+                return text[start : end + 1]
+            return text
+        except Exception:
+            return text
+
     def __retry_unit(self, prepare_prompt_outputs, call_llm_inputs, retry_data: RetryData):
         call_llm = CallLLM(call_llm_inputs)
         call_llm_outputs = call_llm.run()
         self.__record_status_or_raise(retry_data, call_llm)
 
         if self.is_json_mode:
             json_responses = []
+
             for response in call_llm_outputs.get("openai_responses"):
                 try:
+                    # For models that don't support JSON mode, extract JSON from the text response first
+                    if self.is_json_mode_unsupported:
+                        response = self.__extract_json_from_text(response)
+
                     json_response = self.__json_loads(response)
                     json_responses.append(json_response)
                 except json.JSONDecodeError as e:
@@ -91,6 +114,14 @@ def run(self) -> dict:
         prompts = [dict(role="user", content=self.user)]
         if self.system:
             prompts.insert(0, dict(role="system", content=self.system))
+
+        # Special handling for models that don't support JSON mode
+        if self.is_json_mode_unsupported and self.is_json_mode and self.json_example:
+            # Append JSON example to user message
+            prompts[-1][
+                "content"
+            ] += f"\nPlease format your response as a JSON object like this example:\n{json.dumps(self.json_example, indent=2)}"
+
         prepare_prompt_inputs = dict(
             prompt_template=prompts,
             prompt_values=self.prompt_values,
@@ -100,9 +131,14 @@ def run(self) -> dict:
         self.set_status(prepare_prompt.status, prepare_prompt.status_message)
 
         model_keys = [key for key in self.inputs.keys() if key.startswith("model_")]
-        response_format = dict(type="json_object" if self.is_json_mode else "text")
-        if self.json_example is not None:
-            response_format = example_json_to_schema(self.json_example)
+
+        # Set response format based on model and mode
+        response_format = None
+        if not self.is_json_mode_unsupported:
+            response_format = dict(type="json_object" if self.is_json_mode else "text")
+            if self.json_example is not None:
+                response_format = example_json_to_schema(self.json_example)
+
         call_llm_inputs = {
             "prompts": prepare_prompt_outputs.get("prompts"),
             **{

pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "patchwork-cli"
-version = "0.0.94"
+version = "0.0.95"
 description = ""
 authors = ["patched.codes"]
 license = "AGPL"