lang instruction to beginning of prompt (#233)

* lang instruction to beginning of prompt

* pipecat - support openai compatible llms

Author: rootflo-hardik

wavefront/server/apps/call_processing/call_processing/services/language_detection_tool.py

@@ -150,7 +150,7 @@ async def detect_and_switch_language(params: FunctionCallParams):
                         return
 
                     # Append new language instruction to clean base prompt
-                    updated_content = f'{base_prompt}\n\n{language_instruction}'
+                    updated_content = f'{language_instruction}\n\n{base_prompt}'
                     updated_system_message = {
                         'role': 'system',
                         'content': updated_content,

wavefront/server/apps/call_processing/call_processing/services/llm_service.py

@@ -50,7 +50,13 @@ def create_llm_service(llm_config: Dict[str, Any]):
         logger.info(f'Creating LLM service: {llm_type} / {model}')
 
         if llm_type == 'openai':
-            return LLMServiceFactory._create_openai_llm(api_key, model, parameters)
+            base_url = llm_config.get('base_url')
+            if not base_url or base_url == 'https://api.openai.com/v1':
+                return LLMServiceFactory._create_openai_llm(api_key, model, parameters)
+            else:
+                return LLMServiceFactory._create_openai_compatible_llm(
+                    api_key, model, parameters, base_url
+                )
         elif llm_type == 'azure_openai':
             base_url = llm_config.get('base_url')
             if not base_url:
@@ -97,6 +103,36 @@ def _create_openai_llm(api_key: str, model: str, parameters: Dict[str, Any]):
 
         return OpenAILLMService(api_key=api_key, model=model, params=input_params)
 
+    @staticmethod
+    def _create_openai_compatible_llm(
+        api_key: str, model: str, parameters: Dict[str, Any], base_url: str
+    ):
+        """Create a BaseOpenAILLMService for OpenAI-compatible endpoints"""
+        params_dict = {}
+
+        if 'temperature' in parameters:
+            params_dict['temperature'] = parameters['temperature']
+        if 'max_completion_tokens' in parameters:
+            params_dict['max_completion_tokens'] = parameters['max_completion_tokens']
+        if 'top_p' in parameters:
+            params_dict['top_p'] = parameters['top_p']
+        if 'frequency_penalty' in parameters:
+            params_dict['frequency_penalty'] = parameters['frequency_penalty']
+        if 'presence_penalty' in parameters:
+            params_dict['presence_penalty'] = parameters['presence_penalty']
+        if 'seed' in parameters:
+            params_dict['seed'] = parameters['seed']
+
+        input_params = BaseOpenAILLMService.InputParams(**params_dict)
+
+        logger.info(
+            f"OpenAI-compatible LLM config: model={model}, base_url={base_url}, temp={params_dict.get('temperature', 'default')}"
+        )
+
+        return BaseOpenAILLMService(
+            api_key=api_key, model=model, base_url=base_url, params=input_params
+        )
+
     @staticmethod
     def _create_google_llm(api_key: str, model: str, parameters: Dict[str, Any]):
         """Create Google LLM service"""

wavefront/server/apps/call_processing/call_processing/services/pipecat_service.py

@@ -321,7 +321,7 @@ async def run_conversation(
                 f'- Do NOT switch based on a greeting (e.g. "Namaste") or incidental use of another language.\n'
                 f'- If the user requests a language not in the supported list, apologise and tell them which languages are available. Do not call the switch tool.'
             )
-            system_content = f'{base_system_prompt}\n\n{initial_language_instruction}{language_switching_rules}'
+            system_content = f'{initial_language_instruction}\n\n{base_system_prompt}{language_switching_rules}'
             # Store base prompt without language instruction for switching (rules persist across switches)
             language_state['original_system_prompt'] = (
                 base_system_prompt + language_switching_rules