Upd concise prompt. Upd TTS loader

pipeline.py

@@ -706,7 +706,7 @@ def stream_chat(
     
     medical_model_obj, medical_tokenizer = initialize_medical_model(medical_model)
     
-    base_system_prompt = system_prompt if system_prompt else "As a medical specialist, provide clinical and concise answers. Use Markdown format with bullet points. Do not use tables."
+    base_system_prompt = system_prompt if system_prompt else "As a medical specialist, provide clinical and concise answers. Use Markdown format with bullet points. Do not use tables. Provide answers directly without conversational prefixes like 'Here is...', 'This is...'. Start with the actual content immediately."
     
     context_sections = []
     if clinical_intake_context_block:

search.py

@@ -208,7 +208,7 @@ Search Results:
 {combined_content}
 Summary:"""
     
-    system_prompt = "You are a medical information summarizer. Extract and summarize key medical facts accurately."
+    system_prompt = "You are a medical information summarizer. Extract and summarize key medical facts accurately. Provide the summary directly without conversational prefixes like 'Here is...', 'This is...', or 'To summarize...'. Start with the actual content immediately."
     
     result = await call_agent(
         user_prompt=user_prompt,

supervisor.py

@@ -739,7 +739,7 @@ Your task:
 
 Return the final synthesized answer in Markdown format. Do not add meta-commentary or explanations - just provide the final answer."""
     
-    system_prompt = "You are a medical answer synthesis supervisor. Create comprehensive, well-structured final answers from multiple specialist responses."
+    system_prompt = "You are a medical answer synthesis supervisor. Create comprehensive, well-structured final answers from multiple specialist responses. Provide the answer directly without conversational prefixes like 'Here is...', 'This is...'. Start with the actual content immediately."
     
     result = await call_agent(
         user_prompt=prompt,
@@ -861,7 +861,7 @@ Create an enhanced version of the answer that:
 
 Return the enhanced answer in Markdown format. Do not add meta-commentary."""
     
-    system_prompt = "You are a medical answer enhancement supervisor. Improve answers based on evaluation feedback while maintaining accuracy."
+    system_prompt = "You are a medical answer enhancement supervisor. Improve answers based on evaluation feedback while maintaining accuracy. Provide the enhanced answer directly without conversational prefixes like 'Here is...', 'This is...'. Start with the actual content immediately."
     
     result = await call_agent(
         user_prompt=prompt,

ui.py

@@ -217,7 +217,7 @@ def create_demo():
                     )
                     
                     system_prompt = gr.Textbox(
-                        value="As a medical specialist, provide detailed and accurate answers based on the provided medical documents and context. Ensure all information is clinically accurate and cite sources when available.",
+                        value="As a medical specialist, provide detailed and accurate answers based on the provided medical documents and context. Ensure all information is clinically accurate and cite sources when available. Provide answers directly without conversational prefixes like 'Here is...', 'This is...', or 'To answer your question...'. Start with the actual content immediately.",
                         label="System Prompt",
                         lines=3
                     )

voice.py

@@ -408,13 +408,42 @@ def _generate_speech_via_mcp(text: str):
         logger.warning(f"MCP TTS error (sync wrapper): {e}")
     return None
 
+def _generate_speech_with_gpu(text: str):
+    """Internal GPU-decorated function for TTS generation when TTS is available."""
+    if config.global_tts_model is None:
+        logger.info("[TTS] TTS model not loaded, initializing...")
+        initialize_tts_model()
+    
+    if config.global_tts_model is None:
+        logger.error("[TTS] TTS model not available. Please check dependencies.")
+        return None
+    
+    try:
+        logger.info("[TTS] Running TTS generation...")
+        wav = config.global_tts_model.tts(text)
+        with tempfile.NamedTemporaryFile(delete=False, suffix=".wav") as tmp_file:
+            sf.write(tmp_file.name, wav, samplerate=22050)
+            logger.info(f"[TTS] ✅ Speech generated successfully: {tmp_file.name}")
+            return tmp_file.name
+    except Exception as e:
+        logger.error(f"[TTS] TTS error (local maya1): {e}")
+        import traceback
+        logger.debug(f"[TTS] Full traceback: {traceback.format_exc()}")
+        return None
+
 @spaces.GPU(max_duration=120)
+def _generate_speech_gpu_wrapper(text: str):
+    """GPU wrapper for TTS generation - only called when TTS is available."""
+    return _generate_speech_with_gpu(text)
+
 def generate_speech(text: str):
     """Generate speech from text using local maya1 TTS model (with MCP fallback).
     
     The primary path uses the local TTS model (maya-research/maya1). MCP-based
     TTS is only used as a last-resort fallback if the local model is unavailable
     or fails.
+    
+    This function checks TTS availability before attempting GPU allocation.
     """
     if not text or len(text.strip()) == 0:
         logger.warning("[TTS] Empty text provided")
@@ -422,29 +451,32 @@ def generate_speech(text: str):
     
     logger.info(f"[TTS] Generating speech for text: {text[:50]}...")
     
+    # Check TTS availability first - avoid GPU allocation if not available
     if not TTS_AVAILABLE:
-        logger.error("[TTS] TTS library not installed. Please install TTS to use voice generation.")
-        # As a last resort, try MCP-based TTS if available
-        return _generate_speech_via_mcp(text)
-    
-    if config.global_tts_model is None:
-        logger.info("[TTS] TTS model not loaded, initializing...")
-        initialize_tts_model()
-    
-    if config.global_tts_model is None:
-        logger.error("[TTS] TTS model not available. Please check dependencies.")
-        return _generate_speech_via_mcp(text)
+        logger.warning("[TTS] TTS library not installed. Trying MCP fallback...")
+        # Try MCP-based TTS if available (doesn't require GPU)
+        audio_path = _generate_speech_via_mcp(text)
+        if audio_path:
+            logger.info(f"[TTS] ✅ Generated via MCP fallback: {audio_path}")
+            return audio_path
+        else:
+            logger.error("[TTS] ❌ TTS library not installed and MCP fallback failed. Please install TTS: pip install TTS --no-deps && pip install coqui-tts")
+            return None
     
+    # TTS is available - use GPU-decorated function
     try:
-        logger.info("[TTS] Running TTS generation...")
-        wav = config.global_tts_model.tts(text)
-        with tempfile.NamedTemporaryFile(delete=False, suffix=".wav") as tmp_file:
-            sf.write(tmp_file.name, wav, samplerate=22050)
-            logger.info(f"[TTS] ✅ Speech generated successfully: {tmp_file.name}")
-            return tmp_file.name
+        audio_path = _generate_speech_gpu_wrapper(text)
+        if audio_path:
+            return audio_path
+        else:
+            # GPU generation failed, try MCP fallback
+            logger.warning("[TTS] Local TTS generation failed, trying MCP fallback...")
+            return _generate_speech_via_mcp(text)
     except Exception as e:
-        logger.error(f"[TTS] TTS error (local maya1): {e}")
+        logger.error(f"[TTS] GPU TTS generation error: {e}")
         import traceback
         logger.debug(f"[TTS] Full traceback: {traceback.format_exc()}")
+        # Try MCP fallback on error
+        logger.info("[TTS] Attempting MCP fallback after error...")
         return _generate_speech_via_mcp(text)