Temp

src/llm/processor.py

@@ -13,20 +13,20 @@ class LLMProcessor:
         self,
         api_key: Optional[str] = None,
         base_url: Optional[str] = None,
-        model: str = "gpt-4o",
+        model: Optional[str] = None,
     ):
         """
         Initializes the LLMProcessor.
 
         :param api_key: OpenAI API key. If None, it looks for OPENAI_API_KEY in environment variables.
         :param base_url: OpenAI-compatible base URL (e.g., for vLLM).
-        :param model: The model to use for processing.
+        :param model: The model to use for processing. If None, it looks for LLM_MODEL in environment variables.
         """
         self.client = OpenAI(
             api_key=api_key or os.environ.get("OPENAI_API_KEY"),
             base_url=base_url or os.environ.get("OPENAI_BASE_URL"),
         )
-        self.model = model
+        self.model = model or os.environ.get("LLM_MODEL", "gpt-4o")
 
     def _call_llm(
         self,
@@ -45,6 +45,7 @@ class LLMProcessor:
                     {"role": "user", "content": user_prompt},
                 ],
                 response_format=response_format,
+                extra_body={"include_reasoning": False},
             )
             return response.choices[0].message.content
         except Exception as e:
@@ -55,27 +56,36 @@ class LLMProcessor:
         """
         Stage 1: Raw Transcript -> Filtered Text.
         """
-        return self._call_llm(NOISE_FILTER_SYSTEM_PROMPT, text)
+        result = self._call_llm(NOISE_FILTER_SYSTEM_PROMPT, text)
+        print(f"LLM Processor (Filter): {text} -> {result}")
+        return result
 
     def extract_structured_data(self, filtered_text: str) -> ExtractionResult:
         """
         Stage 2: Filtered Text -> Structured Data.
         """
-        # We use OpenAI's structured output (JSON mode/tool calling) via Pydantic's response_format.
-        # For models that support it, we can pass the Pydantic model directly.
-        # If we are using an older model or vLLM, we might need to manually parse the JSON.
-
-        # Using the newer 'beta.chat.completions.parse' for Pydantic support
+        print(f"LLM Processor (Extract): Calling extraction for: {filtered_text}")
         try:
-            completion = self.client.beta.chat.completions.parse(
+            # Using standard chat.completions.create with JSON mode for better compatibility with vLLM
+            response = self.client.chat.completions.create(
                 model=self.model,
                 messages=[
                     {"role": "system", "content": EXTRACTION_SYSTEM_PROMPT},
                     {"role": "user", "content": filtered_text},
                 ],
-                response_format=ExtractionResult,
+                response_format={"type": "json_object"},
+                extra_body={"include_reasoning": False},
             )
-            return completion.choices[0].message.parsed
+
+            import json
+
+            content = response.choices[0].message.content
+            print(f"LLM Processor (Extract): Raw JSON response: {content}")
+            data = json.loads(content)
+
+            # Map the JSON data to the Pydantic model
+            return ExtractionResult(**data)
+
         except Exception as e:
             print(f"Extraction Error: {e}")
             # Return an empty ExtractionResult if parsing fails