Update app.py

app.py

@@ -1,180 +1,215 @@
 import gradio as gr
-from huggingface_hub import InferenceClient # Keep for direct use if needed, though agent will use its own model
+from huggingface_hub import InferenceClient
 import os
 import json
 import base64
 from PIL import Image
 import io
 
-# Smolagents imports
-from smolagents import CodeAgent, Tool
-from smolagents.models import InferenceClientModel as SmolInferenceClientModel
-# We'll use PIL.Image directly for opening, AgentImage is for agent's internal typing if needed by a tool
-from smolagents.gradio_ui import pull_messages_from_step # For formatting agent steps
-from smolagents.memory import ActionStep, FinalAnswerStep, PlanningStep, MemoryStep # For type checking steps
-from smolagents.models import ChatMessageStreamDelta # For type checking stream deltas
-
-
 ACCESS_TOKEN = os.getenv("HF_TOKEN")
 print("Access token loaded.")
 
-# Function to encode image to base64 (remains useful if we ever need to pass base64 to a non-smolagent component)
-def encode_image(image_path_or_pil):
-    if not image_path_or_pil:
-        print("No image path or PIL Image provided")
+# Function to encode image to base64
+def encode_image(image_path):
+    if not image_path:
+        print("No image path provided")
         return None
     
     try:
-        # print(f"Encoding image: {type(image_path_or_pil)}") # Debug
+        print(f"Encoding image from path: {image_path}")
         
-        if isinstance(image_path_or_pil, Image.Image):
-            image = image_path_or_pil
-        else: # Assuming it's a path
-            image = Image.open(image_path_or_pil)
+        # If it's already a PIL Image
+        if isinstance(image_path, Image.Image):
+            image = image_path
+        else:
+            # Try to open the image file
+            image = Image.open(image_path)
         
+        # Convert to RGB if image has an alpha channel (RGBA)
         if image.mode == 'RGBA':
             image = image.convert('RGB')
         
+        # Encode to base64
         buffered = io.BytesIO()
-        image.save(buffered, format="JPEG") # JPEG is generally smaller for transfer
+        image.save(buffered, format="JPEG")
         img_str = base64.b64encode(buffered.getvalue()).decode("utf-8")
-        # print("Image encoded successfully") # Debug
+        print("Image encoded successfully")
         return img_str
     except Exception as e:
         print(f"Error encoding image: {e}")
         return None
 
-# This function will now set up and run the smolagent
 def respond(
-    message_text, # Text from MultimodalTextbox
-    image_file_paths,  # List of file paths from MultimodalTextbox
-    gradio_history: list[tuple[str, str]], # Gradio history (for context if needed, agent is stateless per call here)
-    system_message_for_agent, # System prompt for the main LLM agent
+    message,
+    image_files,  # Changed parameter name and structure
+    history: list[tuple[str, str]],
+    system_message,
     max_tokens,
     temperature,
     top_p,
     frequency_penalty,
     seed,
-    provider_for_agent_llm, 
-    api_key_for_agent_llm, 
-    model_id_for_agent_llm,
-    model_search_term, # Unused directly by agent logic
-    selected_model_for_agent_llm # Fallback model ID
+    provider,
+    custom_api_key,
+    custom_model,    
+    model_search_term,
+    selected_model
 ):
-    print(f"Respond function called. Message: '{message_text}', Images: {image_file_paths}")
+    print(f"Received message: {message}")
+    print(f"Received {len(image_files) if image_files else 0} images")
+    print(f"History: {history}")
+    print(f"System message: {system_message}")
+    print(f"Max tokens: {max_tokens}, Temperature: {temperature}, Top-P: {top_p}")
+    print(f"Frequency Penalty: {frequency_penalty}, Seed: {seed}")
+    print(f"Selected provider: {provider}")         
+    print(f"Custom API Key provided: {bool(custom_api_key.strip())}")
+    print(f"Selected model (custom_model): {custom_model}")  
+    print(f"Model search term: {model_search_term}")
+    print(f"Selected model from radio: {selected_model}")
+
+    # Determine which token to use
+    token_to_use = custom_api_key if custom_api_key.strip() != "" else ACCESS_TOKEN
+    
+    if custom_api_key.strip() != "":
+        print("USING CUSTOM API KEY: BYOK token provided by user is being used for authentication")
+    else:
+        print("USING DEFAULT API KEY: Environment variable HF_TOKEN is being used for authentication")
+    
+    # Initialize the Inference Client with the provider and appropriate token
+    client = InferenceClient(token=token_to_use, provider=provider)
+    print(f"Hugging Face Inference Client initialized with {provider} provider.")
+
+    # Convert seed to None if -1 (meaning random)
+    if seed == -1:
+        seed = None
+
+    # Create multimodal content if images are present
+    if image_files and len(image_files) > 0:
+        # Process the user message to include images
+        user_content = []
+        
+        # Add text part if there is any
+        if message and message.strip():
+            user_content.append({
+                "type": "text",
+                "text": message
+            })
+        
+        # Add image parts
+        for img in image_files:
+            if img is not None:
+                # Get raw image data from path
+                try:
+                    encoded_image = encode_image(img)
+                    if encoded_image:
+                        user_content.append({
+                            "type": "image_url",
+                            "image_url": {
+                                "url": f"data:image/jpeg;base64,{encoded_image}"
+                            }
+                        })
+                except Exception as e:
+                    print(f"Error encoding image: {e}")
+    else:
+        # Text-only message
+        user_content = message
+
+    # Prepare messages in the format expected by the API
+    messages = [{"role": "system", "content": system_message}]
+    print("Initial messages array constructed.")
+
+    # Add conversation history to the context
+    for val in history:
+        user_part = val[0]
+        assistant_part = val[1]
+        if user_part:
+            # Handle both text-only and multimodal messages in history
+            if isinstance(user_part, tuple) and len(user_part) == 2:
+                # This is a multimodal message with text and images
+                history_content = []
+                if user_part[0]:  # Text
+                    history_content.append({
+                        "type": "text",
+                        "text": user_part[0]
+                    })
+                
+                for img in user_part[1]:  # Images
+                    if img:
+                        try:
+                            encoded_img = encode_image(img)
+                            if encoded_img:
+                                history_content.append({
+                                    "type": "image_url",
+                                    "image_url": {
+                                        "url": f"data:image/jpeg;base64,{encoded_img}"
+                                    }
+                                })
+                        except Exception as e:
+                            print(f"Error encoding history image: {e}")
+                
+                messages.append({"role": "user", "content": history_content})
+            else:
+                # Regular text message
+                messages.append({"role": "user", "content": user_part})
+            print(f"Added user message to context (type: {type(user_part)})")
+        
+        if assistant_part:
+            messages.append({"role": "assistant", "content": assistant_part})
+            print(f"Added assistant message to context: {assistant_part}")
+
+    # Append the latest user message
+    messages.append({"role": "user", "content": user_content})
+    print(f"Latest user message appended (content type: {type(user_content)})")
 
-    token_to_use = api_key_for_agent_llm if api_key_for_agent_llm.strip() != "" else ACCESS_TOKEN
-    model_to_use = model_id_for_agent_llm.strip() if model_id_for_agent_llm.strip() != "" else selected_model_for_agent_llm
+    # Determine which model to use, prioritizing custom_model if provided
+    model_to_use = custom_model.strip() if custom_model.strip() != "" else selected_model
+    print(f"Model selected for inference: {model_to_use}")
 
-    # --- Initialize the LLM for the CodeAgent ---
-    agent_llm_params = {
-        "model_id": model_to_use,
-        "token": token_to_use,
-        # smolagents's InferenceClientModel uses max_tokens for max_new_tokens
+    # Start with an empty string to build the response as tokens stream in
+    response = ""
+    print(f"Sending request to {provider} provider.")
+
+    # Prepare parameters for the chat completion request
+    parameters = {
         "max_tokens": max_tokens,
-        "temperature": temperature if temperature > 0.01 else None, # Some models require temp > 0
-        "top_p": top_p if top_p < 1.0 else None, # Often 1.0 means no top_p
-        "seed": seed if seed != -1 else None,
+        "temperature": temperature,
+        "top_p": top_p,
+        "frequency_penalty": frequency_penalty,
     }
-    if provider_for_agent_llm and provider_for_agent_llm != "hf-inference":
-        agent_llm_params["provider"] = provider_for_agent_llm
     
-    # HFIC specific params, add if not default and supported
-    if frequency_penalty != 0.0:
-         agent_llm_params["frequency_penalty"] = frequency_penalty
-         
-    agent_llm = SmolInferenceClientModel(**agent_llm_params)
-    print(f"Smolagents LLM for agent initialized: model='{model_to_use}', provider='{provider_for_agent_llm or 'default'}'")
+    if seed is not None:
+        parameters["seed"] = seed
 
-    # --- Define Tools for the Agent ---
-    agent_tools = []
+    # Use the InferenceClient for making the request
     try:
-        image_gen_tool = Tool.from_space(
-            space_id="black-forest-labs/FLUX.1-schnell",
-            name="image_generator",
-            description="Generates an image from a textual prompt. Input is a single string argument named 'prompt'. Output is an image file path.",
-            token=token_to_use 
+        # Create a generator for the streaming response
+        stream = client.chat_completion(
+            model=model_to_use,
+            messages=messages,
+            stream=True,
+            **parameters
         )
-        agent_tools.append(image_gen_tool)
-        print("Image generation tool loaded: black-forest-labs/FLUX.1-schnell")
-    except Exception as e:
-        print(f"Error loading image generation tool: {e}")
-        yield f"Error: Could not load image generation tool. {e}"
-        return
-
-    # --- Initialize the CodeAgent ---
-    # If system_message_for_agent is empty, CodeAgent will use its default.
-    # The default is usually good as it explains how to use tools.
-    agent = CodeAgent(
-        tools=agent_tools,
-        model=agent_llm,
-        system_prompt=system_message_for_agent if system_message_for_agent and system_message_for_agent.strip() else None,
-        # add_base_tools=True, # Consider adding Python interpreter, etc.
-        stream_outputs=True # Important for Gradio streaming
-    )
-    print("Smolagents CodeAgent initialized.")
-
-    # --- Prepare task and image inputs for the agent ---
-    agent_task_text = message_text
-    
-    pil_images_for_agent = []
-    if image_file_paths:
-        for file_path in image_file_paths:
-            try:
-                pil_images_for_agent.append(Image.open(file_path))
-            except Exception as e:
-                print(f"Error opening image file {file_path} for agent: {e}")
-    
-    print(f"Agent task: '{agent_task_text}'")
-    if pil_images_for_agent:
-        print(f"Passing {len(pil_images_for_agent)} image(s) to agent.")
-
-    # --- Run the agent and stream response ---
-    # Agent is reset each turn. For conversational memory, agent instance
-    # would need to be stored in session_state and agent.run(..., reset=False) used.
-    
-    current_agent_response_text = ""
-    try:
-        # The agent.run method returns a generator when stream=True
-        for step_item in agent.run(
-            task=agent_task_text, 
-            images=pil_images_for_agent, 
-            stream=True, 
-            reset=True # Explicitly reset for stateless operation per call
-        ):
-            if isinstance(step_item, ChatMessageStreamDelta):
-                if step_item.content:
-                    current_agent_response_text += step_item.content
-                    yield current_agent_response_text # Yield accumulated text
-            
-            elif isinstance(step_item, (ActionStep, PlanningStep, FinalAnswerStep)):
-                # A structured step. Format it for Gradio.
-                # pull_messages_from_step yields gr.ChatMessage objects.
-                for gradio_chat_msg in pull_messages_from_step(step_item, skip_model_outputs=agent.stream_outputs):
-                    # The 'bot' function will handle these gr.ChatMessage objects.
-                    yield gradio_chat_msg # Yield the gr.ChatMessage object directly
-                current_agent_response_text = "" # Reset text buffer after a structured step
-            
-            # else:
-                # print(f"Unhandled stream item type: {type(step_item)}") # Debug
-
-        # If there's any remaining text not part of a gr.ChatMessage, yield it.
-        # This usually shouldn't happen if stream_to_gradio logic is followed,
-        # as text deltas should be part of the last gr.ChatMessage or yielded before it.
-        # However, if the agent's final textual answer comes as pure deltas after all steps.
-        if current_agent_response_text and not isinstance(step_item, FinalAnswerStep):
-             # Check if the last yielded item already contains this text
-            if not (isinstance(step_item, gr.ChatMessage) and step_item.content == current_agent_response_text):
-                 yield current_agent_response_text
-
-
+        
+        print("Received tokens: ", end="", flush=True)
+        
+        # Process the streaming response
+        for chunk in stream:
+            if hasattr(chunk, 'choices') and len(chunk.choices) > 0:
+                # Extract the content from the response
+                if hasattr(chunk.choices[0], 'delta') and hasattr(chunk.choices[0].delta, 'content'):
+                    token_text = chunk.choices[0].delta.content
+                    if token_text:
+                        print(token_text, end="", flush=True)
+                        response += token_text
+                        yield response
+        
+        print()
     except Exception as e:
-        error_message = f"Error during agent execution: {str(e)}"
-        print(error_message)
-        yield error_message # Yield the error message to be displayed in UI
-
-    print("Agent run completed.")
+        print(f"Error during inference: {e}")
+        response += f"\nError: {str(e)}"
+        yield response
 
+    print("Completed response generation.")
 
 # Function to validate provider selection based on BYOK
 def validate_provider(api_key, provider):
@@ -184,15 +219,16 @@ def validate_provider(api_key, provider):
 
 # GRADIO UI
 with gr.Blocks(theme="Nymbo/Nymbo_Theme") as demo:
+    # Create the chatbot component
     chatbot = gr.Chatbot(
         height=600, 
         show_copy_button=True, 
-        placeholder="Select a model and begin chatting. Now uses smolagents with tools!",
-        layout="panel",
-        bubble_full_width=False # For better display of images/files
+        placeholder="Select a model and begin chatting. Now supports multiple inference providers and multimodal inputs",
+        layout="panel"
     )
     print("Chatbot interface created.")
     
+    # Multimodal textbox for messages (combines text and file uploads)
     msg = gr.MultimodalTextbox(
         placeholder="Type a message or upload images...",
         show_label=False,
@@ -203,201 +239,329 @@ with gr.Blocks(theme="Nymbo/Nymbo_Theme") as demo:
         sources=["upload"]
     )
     
+    # Note: We're removing the separate submit button since MultimodalTextbox has its own
+
+    # Create accordion for settings
     with gr.Accordion("Settings", open=False):
+        # System message
         system_message_box = gr.Textbox(
-            value="You are a helpful AI assistant. You can generate images if asked. Be precise with your prompts for image generation.", 
-            placeholder="You are a helpful AI assistant.",
-            label="System Prompt for Agent"
+            value="You are a helpful AI assistant that can understand images and text.", 
+            placeholder="You are a helpful assistant.",
+            label="System Prompt"
         )
         
+        # Generation parameters
         with gr.Row():
             with gr.Column():
-                max_tokens_slider = gr.Slider(minimum=1, maximum=4096, value=1024, step=1, label="Max New Tokens")
-                temperature_slider = gr.Slider(minimum=0.0, maximum=2.0, value=0.7, step=0.01, label="Temperature")
-                top_p_slider = gr.Slider(minimum=0.0, maximum=1.0, value=0.95, step=0.01, label="Top-P")
+                max_tokens_slider = gr.Slider(
+                    minimum=1,
+                    maximum=4096,
+                    value=512,
+                    step=1,
+                    label="Max tokens"
+                )
+                
+                temperature_slider = gr.Slider(
+                    minimum=0.1,
+                    maximum=4.0,
+                    value=0.7,
+                    step=0.1,
+                    label="Temperature"
+                )
+                
+                top_p_slider = gr.Slider(
+                    minimum=0.1,
+                    maximum=1.0,
+                    value=0.95,
+                    step=0.05,
+                    label="Top-P"
+                )
+                
             with gr.Column():
-                frequency_penalty_slider = gr.Slider(minimum=-2.0, maximum=2.0, value=0.0, step=0.1, label="Frequency Penalty")
-                seed_slider = gr.Slider(minimum=-1, maximum=65535, value=-1, step=1, label="Seed (-1 for random)")
+                frequency_penalty_slider = gr.Slider(
+                    minimum=-2.0,
+                    maximum=2.0,
+                    value=0.0,
+                    step=0.1,
+                    label="Frequency Penalty"
+                )
+                
+                seed_slider = gr.Slider(
+                    minimum=-1,
+                    maximum=65535,
+                    value=-1,
+                    step=1,
+                    label="Seed (-1 for random)"
+                )
         
+        # Provider selection
         providers_list = [
-            "hf-inference", "cerebras", "together", "sambanova", "novita", 
-            "cohere", "fireworks-ai", "hyperbolic", "nebius",
+            "hf-inference",  # Default Hugging Face Inference
+            "cerebras",      # Cerebras provider
+            "together",      # Together AI
+            "sambanova",     # SambaNova
+            "novita",        # Novita AI
+            "cohere",        # Cohere
+            "fireworks-ai",  # Fireworks AI
+            "hyperbolic",    # Hyperbolic
+            "nebius",        # Nebius
         ]
-        provider_radio = gr.Radio(choices=providers_list, value="hf-inference", label="Inference Provider for Agent's LLM")
-        byok_textbox = gr.Textbox(value="", label="BYOK (Your HF Token or Provider API Key)", info="Enter API key for the selected provider. Uses HF_TOKEN if empty.", placeholder="Enter your API token", type="password")
-        custom_model_box = gr.Textbox(value="", label="Custom Model ID for Agent's LLM", info="(Optional) Provide a custom model ID. Overrides featured model.", placeholder="meta-llama/Llama-3.3-70B-Instruct")
-        model_search_box = gr.Textbox(label="Filter Featured Models", placeholder="Search for a featured model...", lines=1)
         
+        provider_radio = gr.Radio(
+            choices=providers_list,
+            value="hf-inference",
+            label="Inference Provider",
+        )
+        
+        # New BYOK textbox
+        byok_textbox = gr.Textbox(
+            value="",
+            label="BYOK (Bring Your Own Key)",
+            info="Enter a custom Hugging Face API key here. When empty, only 'hf-inference' provider can be used.",
+            placeholder="Enter your Hugging Face API token",
+            type="password"  # Hide the API key for security
+        )
+        
+        # Custom model box
+        custom_model_box = gr.Textbox(
+            value="",
+            label="Custom Model",
+            info="(Optional) Provide a custom Hugging Face model path. Overrides any selected featured model.",
+            placeholder="meta-llama/Llama-3.3-70B-Instruct"
+        )
+        
+        # Model search
+        model_search_box = gr.Textbox(
+            label="Filter Models",
+            placeholder="Search for a featured model...",
+            lines=1
+        )
+        
+        # Featured models list
+        # Updated to include multimodal models
         models_list = [
-            "meta-llama/Llama-3.3-70B-Instruct", "meta-llama/Llama-3.1-70B-Instruct", "meta-llama/Llama-3.0-70B-Instruct",
-            "meta-llama/Llama-3.2-11B-Vision-Instruct", "meta-llama/Llama-3.2-3B-Instruct", "meta-llama/Llama-3.2-1B-Instruct",
-            "meta-llama/Llama-3.1-8B-Instruct", "NousResearch/Hermes-3-Llama-3.1-8B", "NousResearch/Nous-Hermes-2-Mixtral-8x7B-DPO",
-            "mistralai/Mistral-Nemo-Instruct-2407", "mistralai/Mixtral-8x7B-Instruct-v0.1", "mistralai/Mistral-7B-Instruct-v0.3",
-            "Qwen/Qwen3-235B-A22B", "Qwen/Qwen3-32B", "Qwen/Qwen2.5-72B-Instruct", "Qwen/Qwen2.5-3B-Instruct",
-            "Qwen/Qwen2.5-Coder-32B-Instruct", "microsoft/Phi-3.5-mini-instruct", "microsoft/Phi-3-mini-128k-instruct",
+            "meta-llama/Llama-3.2-11B-Vision-Instruct",
+            "meta-llama/Llama-3.3-70B-Instruct",
+            "meta-llama/Llama-3.1-70B-Instruct",
+            "meta-llama/Llama-3.0-70B-Instruct",
+            "meta-llama/Llama-3.2-3B-Instruct",
+            "meta-llama/Llama-3.2-1B-Instruct",
+            "meta-llama/Llama-3.1-8B-Instruct",
+            "NousResearch/Hermes-3-Llama-3.1-8B",
+            "NousResearch/Nous-Hermes-2-Mixtral-8x7B-DPO",
+            "mistralai/Mistral-Nemo-Instruct-2407",
+            "mistralai/Mixtral-8x7B-Instruct-v0.1",
+            "mistralai/Mistral-7B-Instruct-v0.3",
+            "mistralai/Mistral-7B-Instruct-v0.2",
+            "Qwen/Qwen3-235B-A22B",
+            "Qwen/Qwen3-32B",
+            "Qwen/Qwen2.5-72B-Instruct",
+            "Qwen/Qwen2.5-3B-Instruct",
+            "Qwen/Qwen2.5-0.5B-Instruct",
+            "Qwen/QwQ-32B",
+            "Qwen/Qwen2.5-Coder-32B-Instruct",
+            "microsoft/Phi-3.5-mini-instruct",
+            "microsoft/Phi-3-mini-128k-instruct",
+            "microsoft/Phi-3-mini-4k-instruct",
         ]
-        featured_model_radio = gr.Radio(label="Select a Featured Model for Agent's LLM", choices=models_list, value="meta-llama/Llama-3.3-70B-Instruct", interactive=True)
+
+        featured_model_radio = gr.Radio(
+            label="Select a model below",
+            choices=models_list,
+            value="meta-llama/Llama-3.2-11B-Vision-Instruct",  # Default to a multimodal model
+            interactive=True
+        )
         
         gr.Markdown("[View all Text-to-Text models](https://huggingface.co/models?inference_provider=all&pipeline_tag=text-generation&sort=trending) | [View all multimodal models](https://huggingface.co/models?inference_provider=all&pipeline_tag=image-text-to-text&sort=trending)")
 
-    # Chat history state (using gr.State to manage it properly)
-    # The chatbot's value itself will be the history display.
-    # We might need a separate gr.State if agent needs to be conversational across turns.
-    # For now, agent is stateless per turn.
+    # Chat history state
+    chat_history = gr.State([])
+    
+    # Function to filter models
+    def filter_models(search_term):
+        print(f"Filtering models with search term: {search_term}")
+        filtered = [m for m in models_list if search_term.lower() in m.lower()]
+        print(f"Filtered models: {filtered}")
+        return gr.update(choices=filtered)
+
+    # Function to set custom model from radio
+    def set_custom_model_from_radio(selected):
+        print(f"Featured model selected: {selected}")
+        return selected
 
     # Function for the chat interface
-    def user(user_multimodal_input_dict, history):
-        print(f"User input: {user_multimodal_input_dict}")
-        text_content = user_multimodal_input_dict.get("text", "")
-        files = user_multimodal_input_dict.get("files", [])
+    def user(user_message, history):
+        # Debug logging for troubleshooting
+        print(f"User message received: {user_message}")
         
-        user_display_parts = []
-        if text_content and text_content.strip():
-            user_display_parts.append(text_content)
-        for file_path_obj in files: # file_path_obj is a tempfile._TemporaryFileWrapper
-            user_display_parts.append((file_path_obj.name, os.path.basename(file_path_obj.name)))
-            
-        if not user_display_parts:
+        # Skip if message is empty (no text and no files)
+        if not user_message or (not user_message.get("text") and not user_message.get("files")):
+            print("Empty message, skipping")
+            return history
+        
+        # Prepare multimodal message format
+        text_content = user_message.get("text", "").strip()
+        files = user_message.get("files", [])
+        
+        print(f"Text content: {text_content}")
+        print(f"Files: {files}")
+        
+        # If both text and files are empty, skip
+        if not text_content and not files:
+            print("No content to display")
             return history
+        
+        # Add message with images to history
+        if files and len(files) > 0:
+            # Add text message first if it exists
+            if text_content:
+                # Add a separate text message
+                print(f"Adding text message: {text_content}")
+                history.append([text_content, None])
             
-        # Append the user's multimodal message to history for display
-        # The actual data (dict) is passed to `bot` function separately.
-        history.append([user_display_parts if len(user_display_parts) > 1 else user_display_parts[0], None])
-        return history
-
+            # Then add each image file separately
+            for file_path in files:
+                if file_path and isinstance(file_path, str):
+                    print(f"Adding image: {file_path}")
+                    # Add image as a separate message with no text
+                    history.append([f"![Image]({file_path})", None])
+            
+            return history
+        else:
+            # For text-only messages
+            print(f"Adding text-only message: {text_content}")
+            history.append([text_content, None])
+            return history
+    
+    # Define bot response function
     def bot(history, system_msg, max_tokens, temperature, top_p, freq_penalty, seed, provider, api_key, custom_model, search_term, selected_model):
-        if not history or not history[-1][0]: # If no user input
-            yield history
-            return
-
-        # The user's input (text and list of file paths) is in history[-1][0]
-        # If `user` function stores the dict:
-        raw_user_input_dict = history[-1][0] if isinstance(history[-1][0], dict) else {"text": str(history[-1][0]), "files": []}
+        # Check if history is valid
+        if not history or len(history) == 0:
+            print("No history to process")
+            return history
         
-        # If `user` function stores formatted display parts:
-        # We need to reconstruct or rely on msg input to bot.
-        # For now, assuming msg.submit passes the raw dict.
-        # Let's adjust the Gradio flow to pass `msg` directly to `bot` as well.
-
-        # The `msg` variable in `msg.submit` holds the raw MultimodalTextbox output.
-        # We need to pass this raw dict to `respond`.
-        # The `history` is for display.
+        # Get the most recent message and detect if it's an image
+        user_message = history[-1][0]
+        print(f"Processing user message: {user_message}")
         
-        # This part is tricky as `bot` gets `history` which is already formatted for display.
-        # A common pattern is to pass `msg` (raw input) also to `bot`.
-        # Let's assume `history[-1][0]` contains enough info or we adjust `user` fn.
-        # For simplicity, let's assume `user` stores the raw dict if needed,
-        # or `bot` can parse `history[-1][0]` if it's a string/list of tuples.
-
-        # Let's assume `history[-1][0]` is the raw `user_multimodal_input_dict`
-        # This means the `user` function must append it like: `history.append([user_multimodal_input_dict, None])`
-        # And the chatbot will display `str(user_multimodal_input_dict)`.
-        # This is what the current `user` function does.
-
-        user_input_data = history[-1][0] # This should be the dict from MultimodalTextbox
-        text_input_for_agent = user_input_data.get("text", "")
-        # Files from MultimodalTextbox are temp file paths
-        image_file_paths_for_agent = [f.name for f in user_input_data.get("files", []) if hasattr(f, 'name')]
-
-
-        history[-1][1] = "" # Initialize assistant's part for streaming
+        is_image = False
+        image_path = None
+        text_content = user_message
         
-        # Buffer for current text stream from agent
-        # Handles both pure text deltas and text content from gr.ChatMessage
-        current_text_for_turn = ""
-
-        for item in respond(
-            message_text=text_input_for_agent,
-            image_file_paths=image_file_paths_for_agent,
-            gradio_history=history[:-1], # Pass previous turns for context if agent uses it
-            system_message_for_agent=system_msg,
-            max_tokens=max_tokens, temperature=temperature, top_p=top_p,
-            frequency_penalty=freq_penalty, seed=seed,
-            provider_for_agent_llm=provider, api_key_for_agent_llm=api_key,
-            model_id_for_agent_llm=custom_model,
-            model_search_term=search_term, # unused
-            selected_model_for_agent_llm=selected_model
-        ):
-            if isinstance(item, str): # LLM text delta from agent's thought or textual answer
-                current_text_for_turn = item 
-                history[-1][1] = current_text_for_turn
-            elif isinstance(item, gr.ChatMessage):
-                # This is a structured step (thought, tool output, image, etc.)
-                # We need to append this to the history as a new message or part of current message.
-                # For simplicity, let's append its string content to the current turn's assistant message.
-                # If it's an image/file, we'll represent it as a markdown link.
-                if isinstance(item.content, str):
-                    current_text_for_turn = item.content # Replace if it's a full message
-                elif isinstance(item.content, dict) and "path" in item.content:
-                    # This is typically an image or audio file
-                    file_path = item.content["path"]
-                    # We need to make this file accessible to Gradio if it's temporary from agent
-                    # For now, just put a placeholder.
-                    # If it's an output from a tool, the path might be relative to where smolagents saves it.
-                    # Gradio needs an absolute path or a URL.
-                    # A common pattern is to copy temp files to a static dir served by Gradio or use gr.File.
-                    # For now, let's assume Gradio can handle local paths if they are in a folder it knows.
-                    # We'll display it as a tuple for Gradio Chatbot.
-                    # This means history[-1][1] needs to become a list.
-                    
-                    # If current_text_for_turn is not empty, make history[-1][1] a list
-                    if current_text_for_turn and not isinstance(history[-1][1], list):
-                        history[-1][1] = [current_text_for_turn]
-                    elif not current_text_for_turn and not isinstance(history[-1][1], list):
-                         history[-1][1] = []
-
-
-                    alt_text = item.metadata.get("title", os.path.basename(file_path)) if item.metadata else os.path.basename(file_path)
-                    
-                    # Add as new component to the list for current assistant message
-                    if isinstance(history[-1][1], list):
-                        history[-1][1].append((file_path, alt_text))
-                    else: # Should have been made a list above
-                        history[-1][1] = [(file_path, alt_text)]
-                    
-                    current_text_for_turn = "" # Reset text buffer after a file
-                
-                # If it's not a delta, but a full message, replace the current text
-                if not isinstance(history[-1][1], list): # if it hasn't become a list due to file
-                    history[-1][1] = current_text_for_turn
-
-            yield history
-
-    # Event handlers
-    # `msg.submit`'s first argument is the function to call.
-    # Its `inputs` are the Gradio components whose values are passed to the function.
-    # Its `outputs` are the Gradio components that are updated by the function's return value.
-    # The `user` function now appends the raw dict from MultimodalTextbox to history.
-    # The `bot` function takes this history.
-    
-    # When msg is submitted:
-    # 1. Call `user` to update history with user's input. Output is `chatbot`.
-    # 2. Then call `bot` with the updated history. Output is `chatbot`.
-    # 3. Then clear `msg`
+        # Check if this is an image message (marked with ![Image])
+        if isinstance(user_message, str) and user_message.startswith("![Image]("):
+            is_image = True
+            # Extract image path from markdown format ![Image](path)
+            image_path = user_message.replace("![Image](", "").replace(")", "")
+            print(f"Image detected: {image_path}")
+            text_content = ""  # No text for image-only messages
+        
+        # Look back for text context if this is an image
+        text_context = ""
+        if is_image and len(history) > 1:
+            # Use the previous message as context if it's text
+            prev_message = history[-2][0]
+            if isinstance(prev_message, str) and not prev_message.startswith("![Image]("):
+                text_context = prev_message
+                print(f"Using text context from previous message: {text_context}")
+        
+        # Process message through respond function
+        history[-1][1] = ""
+        
+        # Use either the image or text for the API
+        if is_image:
+            # For image messages
+            for response in respond(
+                text_context,           # Text context from previous message if any
+                [image_path],           # Current image
+                history[:-1],           # Previous history
+                system_msg,
+                max_tokens,
+                temperature,
+                top_p,
+                freq_penalty,
+                seed,
+                provider,
+                api_key,
+                custom_model,
+                search_term,
+                selected_model
+            ):
+                history[-1][1] = response
+                yield history
+        else:
+            # For text-only messages
+            for response in respond(
+                text_content,           # Text message
+                None,                   # No image
+                history[:-1],           # Previous history
+                system_msg,
+                max_tokens,
+                temperature,
+                top_p,
+                freq_penalty,
+                seed,
+                provider,
+                api_key,
+                custom_model,
+                search_term,
+                selected_model
+            ):
+                history[-1][1] = response
+                yield history
+
+    # Event handlers - only using the MultimodalTextbox's built-in submit functionality
     msg.submit(
         user,
         [msg, chatbot],
-        [chatbot], # `user` returns the new history, updating the chatbot display
+        [chatbot],
         queue=False
     ).then(
         bot,
         [chatbot, system_message_box, max_tokens_slider, temperature_slider, top_p_slider, 
          frequency_penalty_slider, seed_slider, provider_radio, byok_textbox, custom_model_box, 
          model_search_box, featured_model_radio],
-        [chatbot] # `bot` yields history updates, streaming to chatbot
+        [chatbot]
     ).then(
-        lambda: {"text": "", "files": []},  # Clear MultimodalTextbox
+        lambda: {"text": "", "files": []},  # Clear inputs after submission
         None,
         [msg]
     )
     
-    model_search_box.change(fn=filter_models, inputs=model_search_box, outputs=featured_model_radio)
-    featured_model_radio.change(fn=set_custom_model_from_radio, inputs=featured_model_radio, outputs=custom_model_box)
-    byok_textbox.change(fn=validate_provider, inputs=[byok_textbox, provider_radio], outputs=provider_radio)
-    provider_radio.change(fn=validate_provider, inputs=[byok_textbox, provider_radio], outputs=provider_radio)
+    # Connect the model filter to update the radio choices
+    model_search_box.change(
+        fn=filter_models,
+        inputs=model_search_box,
+        outputs=featured_model_radio
+    )
+    print("Model search box change event linked.")
+
+    # Connect the featured model radio to update the custom model box
+    featured_model_radio.change(
+        fn=set_custom_model_from_radio,
+        inputs=featured_model_radio,
+        outputs=custom_model_box
+    )
+    print("Featured model radio button change event linked.")
+    
+    # Connect the BYOK textbox to validate provider selection
+    byok_textbox.change(
+        fn=validate_provider,
+        inputs=[byok_textbox, provider_radio],
+        outputs=provider_radio
+    )
+    print("BYOK textbox change event linked.")
+
+    # Also validate provider when the radio changes to ensure consistency
+    provider_radio.change(
+        fn=validate_provider,
+        inputs=[byok_textbox, provider_radio],
+        outputs=provider_radio
+    )
+    print("Provider radio button change event linked.")
 
 print("Gradio interface initialized.")
 
 if __name__ == "__main__":
     print("Launching the demo application.")
-    demo.launch(show_api=False) # show_api=False for cleaner launch, True for API docs
+    demo.launch(show_api=True)