src/interfaces/gradio_interface.py

import gradio as gr
from typing import Generator, List
import json
import os
from pathlib import Path
from src.agents.manager_agent import manager_agent

# Custom streaming implementation for better compatibility


def get_cache_file_path():
    """Returns the path for the bookmark cache file."""
    data_dir = Path("data")
    data_dir.mkdir(exist_ok=True)
    return str(data_dir / "ai_bookmarks_cache.json")


def load_cache():
    """Loads the bookmark cache from JSON file."""
    cache_file = get_cache_file_path()
    if os.path.exists(cache_file):
        try:
            with open(cache_file, "r", encoding="utf-8") as f:
                return json.load(f)
        except Exception as e:
            print(f"Error loading cache: {e}")
    return {"bookmarks": [], "last_updated": None}


def get_categories_data():
    """Get categorized bookmarks data for display."""
    AI_CATEGORIES = {
        "research_breakthroughs": {
            "name": "Research & Breakthroughs",
            "description": "Novel papers, theoretical advances, new architectures, state-of-the-art results.",
            "icon": "🔬",
        },
        "model_releases": {
            "name": "Model Releases & Updates",
            "description": "Launches of new large-language or vision models, version upgrades, open-source checkpoints.",
            "icon": "🚀",
        },
        "tools_frameworks": {
            "name": "Tools, Frameworks & Platforms",
            "description": "SDKs, libraries, cloud services, developer toolkits, hosting/serving solutions.",
            "icon": "🛠️",
        },
        "applications_industry": {
            "name": "Applications & Industry Use Cases",
            "description": "AI in healthcare, finance, manufacturing, marketing, robotics—real-world deployments.",
            "icon": "🏭",
        },
        "regulation_ethics": {
            "name": "Regulation, Ethics & Policy",
            "description": "Government guidelines, ethical debates, bias/fairness studies, compliance news.",
            "icon": "⚖️",
        },
        "investment_funding": {
            "name": "Investment, Funding & M&A",
            "description": "Venture rounds, strategic investments, acquisitions, startup valuations.",
            "icon": "💰",
        },
        "benchmarks_leaderboards": {
            "name": "Benchmarks & Leaderboards",
            "description": "Performance comparisons, academic/industry challenges, leaderboard standings.",
            "icon": "🏆",
        },
        "community_events": {
            "name": "Community, Events & Education",
            "description": "Conferences, workshops, hackathons, courses, tutorials, webinars.",
            "icon": "🎓",
        },
        "security_privacy": {
            "name": "Security, Privacy & Safety",
            "description": "Adversarial attacks, defensive techniques, data-privacy breakthroughs, AI safety research.",
            "icon": "🔒",
        },
        "market_trends": {
            "name": "Market Trends & Analysis",
            "description": "Adoption rates, market forecasts, analyst reports, surveys on AI usage.",
            "icon": "📈",
        },
    }

    cache = load_cache()
    bookmarks = cache.get("bookmarks", [])

    categories_with_content = {}
    category_counts = {}

    # Initialize categories
    for key, data in AI_CATEGORIES.items():
        categories_with_content[key] = {
            "name": data["name"],
            "description": data["description"],
            "icon": data["icon"],
            "bookmarks": [],
        }
        category_counts[key] = 0

    # Categorize bookmarks
    for bookmark in bookmarks:
        category = bookmark.get("category", "uncategorized")
        if category in categories_with_content:
            categories_with_content[category]["bookmarks"].append(bookmark)
            category_counts[category] += 1

    # Sort categories by bookmark count (descending)
    sorted_categories = sorted(categories_with_content.items(), key=lambda x: len(x[1]["bookmarks"]), reverse=True)

    return sorted_categories, category_counts, len(bookmarks)


def create_categories_interface():
    """Create the categories display interface."""

    def refresh_categories():
        """Refresh and display categories data."""
        try:
            sorted_categories, category_counts, total_bookmarks = get_categories_data()

            if total_bookmarks == 0:
                return "## No bookmarks found in cache\n\nPlease use the Chat tab to load and categorize your bookmarks first."

            # Create the display content
            content = "# 🏷️ AI Bookmarks Categories\n\n"
            content += f"**Total Bookmarks:** {total_bookmarks}\n\n"

            # Display top 10 categories
            for i, (category_key, category_data) in enumerate(sorted_categories[:10], 1):
                icon = category_data["icon"]
                name = category_data["name"]
                description = category_data["description"]
                bookmark_count = len(category_data["bookmarks"])

                content += f"## {i}. {icon} {name}\n"
                content += f"**Count:** {bookmark_count} bookmarks\n"
                content += f"**Description:** {description}\n\n"

                # Show first 5 bookmarks for each category
                if bookmark_count > 0:
                    content += "**Recent Bookmarks:**\n"
                    for j, bookmark in enumerate(category_data["bookmarks"][:5], 1):
                        title = bookmark.get("title", "Untitled")
                        url = bookmark.get("url", "")
                        content += f"{j}. [{title}]({url})\n"

                    if bookmark_count > 5:
                        content += f"   ... and {bookmark_count - 5} more\n"
                    content += "\n"
                else:
                    content += "*No bookmarks in this category yet.*\n\n"

                content += "---\n\n"

            return content

        except Exception as e:
            return f"## Error loading categories\n\n{str(e)}"

    with gr.Blocks() as categories_tab:
        gr.Markdown("# 🏷️ AI Bookmark Categories Dashboard")
        gr.Markdown("View your AI bookmarks organized by the main 10 categories.")

        refresh_btn = gr.Button("🔄 Refresh Categories", variant="primary")
        categories_display = gr.Markdown(refresh_categories())

        refresh_btn.click(fn=refresh_categories, outputs=categories_display)

    return categories_tab


def create_about_interface():
    """Create the about page interface."""

    intro_content = """
    # 🧠 About ReMind - AI and Tech News Assistant

    ## 🎥 Project Demo Video

    [![ReMind Demo Video](https://img.youtube.com/vi/CD0j2dGVycs/maxresdefault.jpg)](https://youtu.be/CD0j2dGVycs)

    **Watch the full project demonstration:** [https://youtu.be/CD0j2dGVycs](https://youtu.be/CD0j2dGVycs)

    ---

    ## Bring your AI knowledge to mind.

    **ReMind** is your intelligent digital memory assistant that helps you rediscover, organize, and make sense of your accumulated AI and technology knowledge. In our information-rich world, we often bookmark valuable resources and receive important newsletters only to forget about them later. This system solves this problem by intelligently categorizing and surfacing your digital discoveries when you need them most.
    """

    about_content = """
## 🎯 What ReMind Does

### 🔖 **Smart Bookmark Management**
- Automatically extracts and caches AI bookmarks from your Chrome "AI ressources" folder
- Provides intelligent search and filtering capabilities by domain, keywords, and categories
- Tracks bookmark statistics and usage patterns
- Focuses specifically on AI and technology resources with comprehensive caching system

### 📧 **Newsletter Email Integration**
- Securely accesses emails from trusted AI news sources (news@alphasignal.ai)
- Searches through AI newsletters and updates with intelligent filtering
- Extracts insights from email-based learning resources
- Provides recent email browsing and content reading capabilities

### 🌐 **Real-time Web Search**
- Performs live web searches for the latest AI and tech developments
- Combines cached knowledge with current information
- Supports up to 6-step search processes for comprehensive research
- Delivers real-time results and analysis

### 🏷️ **Intelligent Categorization**
The system automatically organizes your content into **10 key AI categories**:

1. **🔬 Research & Breakthroughs** - Latest papers, theoretical advances, and novel architectures
2. **🚀 Model Releases & Updates** - New AI models, version upgrades, and open-source checkpoints  
3. **🛠️ Tools, Frameworks & Platforms** - Developer SDKs, libraries, cloud services, and deployment solutions
4. **🏭 Applications & Industry Use Cases** - Real-world AI implementations across healthcare, finance, manufacturing
5. **⚖️ Regulation, Ethics & Policy** - AI governance, ethical considerations, and compliance guidelines
6. **💰 Investment, Funding & M&A** - Market movements, startup funding, and acquisition news
7. **🏆 Benchmarks & Leaderboards** - Performance comparisons, academic challenges, and ranking systems
8. **🎓 Community, Events & Education** - Learning resources, conferences, workshops, and tutorials
9. **🔒 Security, Privacy & Safety** - AI safety research, adversarial attacks, and privacy protection
10. **📈 Market Trends & Analysis** - Industry insights, adoption rates, and market forecasts

### 💬 **Multi-Agent Conversational Interface**
- **Manager Agent**: Routes queries to specialized agents based on context
- **Bookmarks Agent**: Handles Chrome bookmark extraction, caching, and management
- **Gmail Agent**: Manages email access, search, and content reading
- **Categorizer Agent**: Provides intelligent classification with manual override capabilities
- **Web Search Agent**: Performs real-time web research and analysis

---

## 🔧 How It Works

**ReMind** is powered by HuggingFace's **Smolagents**. Here is how it works:

- **🤖 Multi-agent orchestration** - Specialized agents work together seamlessly
- **🧠 Real-time reasoning** - Watch AI agents think through problems step-by-step
- **🔄 Dynamic categorization** - Keyword-based classification with manual recategorization
- **🔍 Semantic search** - Find resources across bookmarks, emails, and web results
- **💾 Local caching** - Efficient JSON-based storage for offline access

---

## 🚀 Getting Started

1. **Update Bookmark Cache**: Import your Chrome "AI ressources" bookmarks into the local cache
2. **Categorize Content**: Automatically classify your AI resources into organized categories
3. **Access Email Updates**: Browse recent newsletters and AI news from trusted sources
4. **Search Across Sources**: Use natural language to find specific resources across all data sources
5. **Stay Current**: Combine cached knowledge with real-time web search for latest developments

---

## 🔒 Privacy & Security

- **Local JSON Caching**: Bookmarks processed and stored in local cache files
- **Selective Email Access**: Only accesses specified trusted email sources with OAuth security
- **Controlled Web Access**: Real-time search with responsible usage limits
- **Transparent Operations**: All agent operations are visible and explainable
- **No Data Sharing**: Personal information processed locally with secure authentication

---

## 💡 Why ReMind?

In the fast-moving world of AI and technology, staying informed while managing information overload is challenging. This system transforms your passive bookmark collection and newsletter subscriptions into an active, intelligent knowledge base that:

- **Surfaces forgotten resources** from your Chrome bookmarks
- **Organizes email newsletters** into actionable intelligence  
- **Combines multiple sources** for comprehensive AI knowledge management
- **Provides real-time updates** through web search integration
- **Learns and adapts** through intelligent categorization and recategorization

---

## 🙏 Acknowledgments

Thanks to **Modal Labs**, **Hugging Face**, **Nebius**, **Anthropic**, **OpenAI**, **Hyperbolic Labs**, **MistralAI**, and **Sambanova.AI** for providing credits that power this demo.

---

*"The smartest AI system is one that helps you manage AI knowledge itself."*

**Welcome to ReMind - where your digital AI knowledge becomes your strategic advantage.**
    """

    with gr.Blocks() as about_tab:
        gr.Markdown(intro_content)

        # Add the AI architecture diagram
        gr.Markdown("## 🏗️ System Architecture")
        gr.Image(
            value="data/ai_diagram.png",
            label="ReMind AI System Architecture",
            show_label=True,
            show_download_button=True,
            height=400,
            width=None,
            interactive=False,
        )
        gr.Markdown("*System architecture showing the multi-agent orchestration and data flow in ReMind*")
        gr.Markdown(about_content)
    return about_tab


def sanitize_content(content):
    """Sanitize content to ensure it's a clean string without complex objects"""
    if isinstance(content, str):
        return content
    elif isinstance(content, dict):
        # If content is a dict, convert to string representation
        return str(content)
    elif hasattr(content, "__dict__"):
        # If it's an object with attributes, convert to string
        return str(content)
    else:
        return str(content)


def validate_message_history(history):
    """Validate and return properly formatted message history"""
    validated = []
    for msg in history:
        if isinstance(msg, dict) and "role" in msg and "content" in msg:
            # Ensure content is a string and properly formatted
            content = sanitize_content(msg["content"])

            # Create a clean message dict to avoid any nesting issues
            clean_msg = {"role": str(msg["role"]), "content": content}
            validated.append(clean_msg)
        else:
            print(f"Warning: Invalid message format detected: {msg}")
    return validated


def chat_with_agent(message: str, history: List) -> Generator[List, None, None]:
    """
    Chat with the agent using custom streaming functionality for real-time thinking display
    """
    try:
        # Convert history to proper format if needed
        if history is None:
            history = []

        # Ensure all history items are properly formatted
        formatted_history = []
        for item in history:
            if isinstance(item, dict):
                # Already a dict, check if it has required keys
                if "role" in item and "content" in item:
                    # Ensure content is a simple string
                    content = sanitize_content(item["content"])
                    formatted_history.append({"role": str(item["role"]), "content": content})
                else:
                    # Skip malformed dict items
                    print(f"Warning: Skipping malformed history item: {item}")
                    continue
            elif hasattr(item, "role") and hasattr(item, "content"):
                # ChatMessage object - convert to dict with string content
                content = sanitize_content(item.content)
                formatted_history.append({"role": str(item.role), "content": content})
            elif isinstance(item, (list, tuple)) and len(item) == 2:
                # Legacy format: [user_message, assistant_message] or (user, assistant)
                # Convert to proper message format
                if isinstance(item[0], str) and isinstance(item[1], str):
                    formatted_history.append({"role": "user", "content": str(item[0])})
                    formatted_history.append({"role": "assistant", "content": str(item[1])})
                else:
                    print(f"Warning: Skipping malformed history item: {item}")
                    continue
            else:
                # Unknown format, skip it
                print(f"Warning: Skipping unknown history format: {type(item)} - {item}")
                continue

        # Reset memory for long conversations to prevent token overflow
        reset_memory = len(formatted_history) > 10  # Reset after 5 user-assistant exchanges

        # Start with user message in history
        new_history = formatted_history.copy()

        # Show initial thinking message with spinner
        thinking_message = {
            "role": "assistant",
            "content": "🔄 **Agent Planning** ⏳\n\n💫 Analyzing your request and creating execution plan...\n\n*Please wait while I process your request...*",
        }
        new_history.append(thinking_message)
        yield validate_message_history(new_history)

        # Run agent with streaming enabled
        try:
            # Use agent.run with stream=True to get step-by-step execution
            agent_stream = manager_agent.run(
                message,
                stream=True,
                reset=reset_memory,
            )

            step_count = 0
            for step in agent_stream:
                step_count += 1

                # Update thinking message with current step info and spinner
                if hasattr(step, "step_number") and hasattr(step, "action"):
                    step_content = "⚡ **Agent Working** 🔄\n\n"
                    step_content += f"🔍 **Step {step.step_number}:** *In Progress...*\n\n"

                    if hasattr(step, "thought") and step.thought:
                        step_content += f"💭 **Thought:** {str(step.thought)}\n\n"

                    if hasattr(step, "action") and step.action:
                        step_content += f"🛠️ **Action:** {str(step.action)}\n\n"

                    if hasattr(step, "observations") and step.observations:
                        obs_text = str(step.observations)[:300]
                        if len(str(step.observations)) > 300:
                            obs_text += "..."
                        step_content += f"👁️ **Observation:** {obs_text}\n\n"

                    step_content += "⏳ *Processing next step...*"

                    # Ensure the content is a clean string
                    thinking_message = {"role": "assistant", "content": str(step_content)}
                    new_history[-1] = thinking_message
                    yield validate_message_history(new_history)

        except Exception as stream_error:
            # If streaming fails, fall back to regular execution
            print(f"Streaming failed: {stream_error}, falling back to regular execution")

            thinking_message = {
                "role": "assistant",
                "content": "⚡ **Agent Working** 🔄\n\n💫 Processing your request using available tools...\n\n⏳ *Please wait...*",
            }
            new_history[-1] = thinking_message
            yield validate_message_history(new_history)

            # Execute without streaming
            result = manager_agent.run(
                message,
                stream=False,
                reset=reset_memory,
            )

            # Show tool usage if available
            tool_usage_content = ""
            if (
                hasattr(manager_agent, "memory")
                and hasattr(manager_agent.memory, "steps")
                and manager_agent.memory.steps
            ):
                try:
                    # Get recent action steps
                    action_steps = [step for step in manager_agent.memory.steps if hasattr(step, "step_number")]
                    recent_steps = action_steps[-3:] if len(action_steps) > 3 else action_steps

                    if recent_steps:
                        tool_details = []
                        for step in recent_steps:
                            if hasattr(step, "step_number"):
                                step_info = f"**Step {step.step_number}**"

                                if hasattr(step, "duration") and step.duration:
                                    step_info += f" ({step.duration:.1f}s)"

                                if hasattr(step, "observations") and step.observations:
                                    obs_text = str(step.observations)[:150]
                                    if len(str(step.observations)) > 150:
                                        obs_text += "..."
                                    step_info += f"\n✅ {obs_text}"

                                if hasattr(step, "error") and step.error:
                                    error_text = str(step.error)[:100]
                                    if len(str(step.error)) > 100:
                                        error_text += "..."
                                    step_info += f"\n❌ {error_text}"

                                tool_details.append(step_info)

                        if tool_details:
                            tool_usage_content = "\n\n".join(tool_details)

                except Exception as e:
                    print(f"Error processing agent steps: {e}")
                    tool_usage_content = "Agent executed actions successfully"

            # Update thinking to show completion
            thinking_message = {
                "role": "assistant",
                "content": "✅ **Agent Complete** 🎉\n\n✅ Request processed successfully\n✅ Response prepared",
            }
            new_history[-1] = thinking_message
            yield validate_message_history(new_history)

            # Add tool usage message if there were tools used
            if tool_usage_content:
                tool_message = {
                    "role": "assistant",
                    "content": f"🛠️ **Tools & Actions Used**\n\n{str(tool_usage_content)}",
                }
                new_history.append(tool_message)
                yield validate_message_history(new_history)

            # Add final response
            final_response = str(result) if result else "I couldn't process your request."
            final_message = {"role": "assistant", "content": str(final_response)}
            new_history.append(final_message)
            yield validate_message_history(new_history)
            return

        # If we get here, streaming worked, so get the final result
        # The streaming should have shown all the steps, now get final answer
        thinking_message = {
            "role": "assistant",
            "content": "✅ **Agent Complete** 🎉\n\n✅ All steps executed\n✅ Preparing final response",
        }
        new_history[-1] = thinking_message
        yield validate_message_history(new_history)

        # Get the final result from the agent memory
        final_response = "Task completed successfully!"
        if hasattr(manager_agent, "memory") and hasattr(manager_agent.memory, "steps") and manager_agent.memory.steps:
            # Get the last step's observations as the final answer
            last_step = manager_agent.memory.steps[-1]
            if hasattr(last_step, "observations") and last_step.observations:
                final_response = str(last_step.observations)

        final_message = {"role": "assistant", "content": str(final_response)}
        new_history.append(final_message)
        yield validate_message_history(new_history)

    except Exception as e:
        # Fallback error handling
        error_message = {
            "role": "assistant",
            "content": f"❌ **System Error:** {str(e)}\n\nPlease try again with a different approach.",
        }
        if "new_history" in locals():
            new_history.append(error_message)
            yield validate_message_history(new_history)
        else:
            # If new_history wasn't initialized, create a minimal valid history
            yield validate_message_history([error_message])


# Create the main chat interface
chat_interface = gr.ChatInterface(
    fn=chat_with_agent,
    type="messages",
    title="🔖 AI and Tech News Assistant - Powered by Smolagents",
    description="""
    ## Your Comprehensive AI Assistant! 🤖
    
    I can help you with:
    
    ### 🔖 **Chrome Bookmarks Management**
    - Search and filter AI resources bookmarks
    - Get bookmark statistics and information  
    - Filter bookmarks by domain
    - Cache and manage Chrome bookmarks data
    
    ### 🏷️ **AI News Categorization**
    - Categorize AI bookmarks into 10 predefined categories
    - Get categorization statistics and insights
    - Search bookmarks by specific categories
    - Manually recategorize bookmarks when needed
    
    ### 📧 **Email Management** 
    - Browse recent emails from trusted senders
    - Search emails by keywords (AI, newsletters, updates, etc.)
    - Read full email content when you need details
    
    ### 🌐 **Web Search**
    - Perform web searches for current information
    - Research topics and gather up-to-date data
    
    ---
    **⏱️ Processing Time Note:** Depending on the type of query, processing can take several seconds or minutes to complete.
    """,
    examples=[
        "📧 Show me my latest 5 emails",
        "🔖 Search my AI bookmarks",
        "🤖 Find emails about AI",
        "🌐 Search for latest AI news",
        "💎 What AI resources do I have?",
        "🐙 Filter bookmarks by GitHub domain",
        "📰 Search for newsletter emails",
        "🏷️ Categorize all my AI bookmarks",
        "📊 Show me categorization statistics",
        "🔬 Get research & breakthrough bookmarks",
        "🚀 Show model releases bookmarks",
        "🛠️ Find tools and frameworks bookmarks",
    ],
    show_progress="hidden",
    cache_examples=False,
)

# Create categories and about interfaces
categories_interface = create_categories_interface()
about_interface = create_about_interface()

# Create tabbed interface
demo = gr.TabbedInterface(
    [about_interface, chat_interface, categories_interface],
    ["ℹ️ About", "💬 Chat Assistant", "🏷️ Categories Dashboard"],
    title="ReMind - Bring your past to mind.",
)