app.py

import gradio as gr
import litellm
import pypdf
import base64
import mimetypes
import os

# ==========================================
# CONFIGURATION & SYSTEM INSTRUCTIONS
# ==========================================
# The "openai/" prefix forces LiteLLM to use your OPENAI_API_BASE secret
MODEL_NAME = "openai/gemini-3-pro-preview"

# Hardcoded backend system instructions
SYSTEM_PROMPT = """You are a brilliant, highly capable AI assistant. 
You carefully analyze any images or text documents provided to you. 
If a user uploads code or text files, read the contents and assist them accordingly."""

# ==========================================
# HELPER FUNCTIONS
# ==========================================
def encode_file_to_base64(file_path):
    """Converts an image file to a base64 data URL expected by Vision LLMs."""
    mime_type, _ = mimetypes.guess_type(file_path)
    if not mime_type:
        mime_type = "image/jpeg"
    with open(file_path, "rb") as file:
        encoded = base64.b64encode(file.read()).decode("utf-8")
    return f"data:{mime_type};base64,{encoded}"

# ==========================================
# CORE CHAT LOGIC
# ==========================================
def predict(message, history):
    # 1. Initialize messages with the hidden System Prompt
    messages = [{"role": "system", "content": SYSTEM_PROMPT}]
    
    # 2. Append conversation history
    for msg in history:
        # To save tokens, we only retain the text portions of previous turns
        if isinstance(msg["content"], str):
             messages.append({"role": msg["role"], "content": msg["content"]})
             
    # 3. Process the current user input (Multimodal Dictionary)
    current_content = []
    
    # Extract text from the chat bar
    user_text = message.get("text", "")
    if user_text:
        current_content.append({"type": "text", "text": user_text})
        
    # Extract and process uploaded files
    for file_path in message.get("files", []):
        mime_type, _ = mimetypes.guess_type(file_path)
        
        # --- A. IMAGE HANDLING ---
        if mime_type and mime_type.startswith("image/"):
            base64_url = encode_file_to_base64(file_path)
            current_content.append({
                "type": "image_url",
                "image_url": {"url": base64_url}
            })

        # --- B. PDF HANDLING ---
        elif file_path.lower().endswith(".pdf"):
            try:
                reader = pypdf.PdfReader(file_path)
                pdf_text = ""
                
                # Loop through every page and extract the text
                for i, page in enumerate(reader.pages):
                    extracted = page.extract_text()
                    if extracted:
                        pdf_text += f"--- Page {i+1} ---\n{extracted}\n\n"
                
                # Package it nicely for the LLM prompt
                doc_text = f"\n\n--- PDF Document: {os.path.basename(file_path)} ---\n{pdf_text}\n--- End PDF ---\n"
                current_content.append({"type": "text", "text": doc_text})
                
            except Exception as e:
                error_msg = f"\n[System: Could not read PDF {os.path.basename(file_path)}. Error: {str(e)}]\n"
                current_content.append({"type": "text", "text": error_msg})

        # --- C. GENERIC TEXT/CODE HANDLING ---
        else:
            try:
                with open(file_path, "r", encoding="utf-8") as f:
                    file_text = f.read()
                doc_text = f"\n\n--- Document: {os.path.basename(file_path)} ---\n{file_text}\n--- End Document ---\n"
                current_content.append({"type": "text", "text": doc_text})
            except Exception:
                current_content.append({"type": "text", "text": f"\n[System: Could not read file: {os.path.basename(file_path)}]\n"})
        
    messages.append({"role": "user", "content": current_content})
    
    # 4. Stream the response from LiteLLM
    try:
        response = litellm.completion(
            model=MODEL_NAME,
            messages=messages,
            stream=True
        )
        partial_message = ""
        for chunk in response:
            # Safely extract streamed text chunks
            delta = chunk.choices[0].delta.content
            if delta:
                partial_message += delta
                yield partial_message
                
    except Exception as e:
        yield f"**An error occurred:** {str(e)}\n\n*Check your API keys and LiteLLM model name.*"

# ==========================================
# USER INTERFACE & SECURITY
# ==========================================
demo = gr.ChatInterface(
    fn=predict,
    multimodal=True,       
    textbox=gr.MultimodalTextbox(
        placeholder="Type a message, or drop an image/document here...",
        # Fixed the missing dot on .pdf below:
        file_types=["image", ".txt", ".csv", ".py", ".md", ".json", ".pdf"], 
        scale=7
    ),
    title="OmniChat",
    description="A minimal, multimodal chat interface routed through LiteLLM.",
    fill_height=True,      
)

if __name__ == "__main__":
    
    # Added the custom HTML from earlier to fix the lame UI
    custom_login_html = """
    <div style="text-align: center; padding: 10px;">
        <h2 style="color: #4F46E5; margin-bottom: 5px; font-family: sans-serif;">🌌 OmniChat Secure Portal</h2>
        <p style="color: #6B7280; font-size: 14px; font-family: sans-serif;">Authorized access only. Please provide your credentials to wake up the LLMs.</p>
    </div>
    """

    def verify_login(username, password):
        users_env = os.environ.get("ALLOWED_USERS", "")
        
        if not users_env:
            return False 
            
        try:
            valid_users = dict(pair.split(":") for pair in users_env.split(","))
        except ValueError:
            return False
            
        return valid_users.get(username) == password

    # Pass the custom HTML into the auth_message parameter
    demo.launch(auth=verify_login, auth_message=custom_login_html)