Update app.py

app.py

@@ -25,8 +25,8 @@ except Exception as e:
 @spaces.GPU()
 def analyze_symptoms(symptom_image: Image.Image, symptoms_text: str):
     """
-    Analyzes user's symptoms using the definitive correct calling convention for
-    multimodal chat models in Hugging Face pipelines.
+    Analyzes user's symptoms using a single, formatted prompt string, which is the
+    most reliable method for this pipeline.
     """
     if not model_loaded:
         return "Error: The AI model could not be loaded. Please check the Space logs."
@@ -36,38 +36,35 @@ def analyze_symptoms(symptom_image: Image.Image, symptoms_text: str):
         return "Please describe your symptoms or upload an image for analysis."
 
     try:
-        # --- DEFINITIVE CHAT-BASED PROMPT LOGIC ---
+        # --- DEFINITIVE PROMPT STRING LOGIC ---
+        # We construct a single string with special tokens that the model understands.
+        # This is more robust than the 'messages' format for this pipeline.
+        
         system_instruction = (
             "You are an expert, empathetic AI medical assistant. "
-            "Analyze the potential medical condition based on the user's input. "
+            "Analyze the potential medical condition based on the following information. "
             "Provide a list of possible conditions, your reasoning, and a clear, "
-            "actionable next-steps plan. Begin your analysis by describing the information "
-            "the user provided."
+            "actionable next-steps plan."
         )
 
-        # The user's content is a list of dictionaries describing the multimodal input.
-        user_content = []
+        prompt_parts = ["<start_of_turn>user"]
         
-        # The text part of the prompt.
-        user_content.append({"type": "text", "text": symptoms_text})
+        # Combine the user's text and the system instruction.
+        full_user_prompt = f"{symptoms_text}\n\n{system_instruction}"
+        prompt_parts.append(full_user_prompt)
 
-        # The images to be passed to the pipeline.
-        images_to_pass = []
+        # The <image> token is a placeholder that tells the model where to "look" at the image.
         if symptom_image:
-            # 1. Add a placeholder to the content list.
-            user_content.append({"type": "image"})
-            # 2. Add the actual PIL image object to a separate list.
-            images_to_pass.append(symptom_image)
-
-        # Construct the final messages list for the chat template.
-        messages = [
-            {"role": "system", "content": system_instruction},
-            {"role": "user", "content": user_content},
-        ]
+            prompt_parts.append("<image>")
+            
+        prompt_parts.append("<start_of_turn>model")
+        
+        # Join all parts into the final prompt string.
+        prompt = "\n".join(prompt_parts)
         
-        print("Generating pipeline output with separate image argument...")
+        print(f"Generating pipeline output with prompt:\n{prompt}")
 
-        # Generation parameters must be in a `generate_kwargs` dictionary.
+        # Generation parameters must be in a 'generate_kwargs' dictionary.
         generate_kwargs = {
             "max_new_tokens": 512,
             "do_sample": True,
@@ -75,34 +72,48 @@ def analyze_symptoms(symptom_image: Image.Image, symptoms_text: str):
         }
         
         # --- THE DEFINITIVE PIPELINE CALL ---
-        # The pipeline call requires three distinct things:
-        # 1. The `messages` list describing the chat structure.
-        # 2. The `images` list containing the actual PIL image data.
-        # 3. The `generate_kwargs` dictionary for generation parameters.
-        output = pipe(
-            messages,
-            images=images_to_pass,
-            generate_kwargs=generate_kwargs
-        )
+        # We pass the prompt string and, if available, the image list.
+        # This calling signature correctly provides the data to the preprocessor.
+        
+        if symptom_image:
+            # Provide both the image data and the text prompt.
+            output = pipe(
+                images=[symptom_image],
+                prompt=prompt,
+                generate_kwargs=generate_kwargs
+            )
+        else:
+            # Provide only the text prompt.
+            output = pipe(
+                prompt=prompt,
+                generate_kwargs=generate_kwargs
+            )
 
         print("Pipeline Output:", output)
 
-        # --- SIMPLIFIED OUTPUT PROCESSING ---
-        # For this pipeline, the output is a list with one dictionary.
-        # The 'generated_text' key contains the AI's response string directly.
+        # --- CORRECTED OUTPUT PROCESSING ---
+        # The output contains the full prompt plus the response. We need to split
+        # the response from the prompt.
         if output and isinstance(output, list) and 'generated_text' in output[0]:
-            result = output[0]['generated_text']
+            full_text = output[0]['generated_text']
+            # The model's response comes after the final '<start_of_turn>model' token.
+            parts = full_text.split("<start_of_turn>model\n")
+            if len(parts) > 1:
+                result = parts[1]
+            else:
+                result = full_text # Fallback in case of unexpected format
         else:
             result = "The model did not return a valid response. Please try again."
 
         disclaimer = "\n\n***Disclaimer: I am an AI assistant and not a medical professional. This is not a diagnosis. Please consult a doctor for any health concerns.***"
         
-        return result + disclaimer
+        return result.strip() + disclaimer
 
     except Exception as e:
         print(f"An exception occurred during analysis: {type(e).__name__}: {e}")
         return f"An error occurred during analysis. Please check the logs for details: {str(e)}"
 
+
 # --- Gradio Interface (No changes needed) ---
 with gr.Blocks(theme=gr.themes.Soft(), title="AI Symptom Analyzer") as demo:
     gr.HTML("""