app.py

import gradio as gr
import pandas as pd
import os
import re
import html
import time
import sys
from pathlib import Path

# Import Groq API client
try:
    from groq import Groq
except ImportError:
    print("Groq API not installed. Run: pip install groq")

# Function to load all CSV files from the current directory
def load_csv_files():
    csv_files = {}
    current_dir = Path(".")
    for file in current_dir.glob("*_sorted.csv"):
        try:
            df = pd.read_csv(file, encoding='utf-8')
            # Fill NaN values with empty strings to avoid issues
            df = df.fillna("")
            # Clean the city name from the filename
            city_name = file.stem.replace('_sorted', '')
            city_name = city_name.replace('_', ' ').title()
            csv_files[city_name] = df
        except Exception as e:
            print(f"Error loading {file}: {e}")
    return csv_files

# Function to get unique queries for a specific city
def get_queries_for_city(city):
    if city not in all_data:
        return []
    
    # Get unique queries from the dataframe
    queries = all_data[city]['query'].dropna().unique().tolist()
    
    # Sort queries and filter out empty strings
    queries = sorted([str(q) for q in queries if q and str(q).strip()])
    
    return queries

# Function to find entries that have empty or missing queries
def find_empty_queries(city, preserve_order=True):
    data = all_data.get(city)
    if data is None:
        return "City data not found"
    
    results = []
    for i, row in data.iterrows():
        # Check if query is empty or NaN
        if pd.isna(row['query']) or str(row['query']).strip() == "":
            # Make sure all values are strings and handle NaN/None values
            context = str(row['context']) if not pd.isna(row['context']) else ""
            query = "(No Query)" if pd.isna(row['query']) else str(row['query'])
            url = str(row['url']) if not pd.isna(row['url']) else ""
            
            results.append({
                'url': url,
                'context': context,
                'query': query,
                'original_index': i  # Store the original row index
            })
    
    # Format results using the same HTML formatting as search_data
    if not results:
        return "No entries without queries found"
    
    # Sort results by their original index if preserve_order is True
    if preserve_order:
        results.sort(key=lambda x: x['original_index'])
    
    # Create HTML formatted results for clickable links with better styling
    formatted_results = "<div class='search-results'>"
    for i, result in enumerate(results, 1):
        url = result['url']
        url_safe = html.escape(url)
        original_idx = result['original_index'] + 1  # +1 for 1-based indexing for display
        
        formatted_results += f"<div class='result-item'>"
        formatted_results += f"<h3>Entry Without Query #{i} <span class='original-index'>(Dataset Row: {original_idx})</span></h3>"
        formatted_results += f"<p><b>URL:</b> <a href='{url_safe}' target='_blank'>{url_safe}</a></p>"
        
        # Handle context display safely
        context = result['context']
        try:
            context_preview = context[:300] + ('...' if len(context) > 300 else '')
            context_preview = html.escape(context_preview)
        except (TypeError, AttributeError):
            context_preview = html.escape(str(context))
            
        formatted_results += f"<p><b>Context:</b> {context_preview}</p>"
        formatted_results += "</div><hr>"
    
    formatted_results += "</div>"
    return formatted_results

# Function to search through the dataframes based on query
def search_data(city, search_type, search_query, case_sensitive=False, preserve_order=True):
    data = all_data.get(city)
    if data is None:
        return "City data not found"
    
    # Check if search_query is empty or None
    if not search_query or str(search_query).strip() == "":
        return "Please enter a search query"
    
    # Ensure search_query is a string
    search_query = str(search_query)
        
    # Convert search query to lowercase if not case sensitive
    if not case_sensitive:
        search_query = search_query.lower()
    
    results = []
    
    if search_type == "Simple Text Search":
        for i, row in data.iterrows():
            # Make sure all values are strings and handle NaN/None values
            context = str(row['context']) if not pd.isna(row['context']) else ""
            query = str(row['query']) if not pd.isna(row['query']) else ""
            url = str(row['url']) if not pd.isna(row['url']) else ""
            
            # Check in context and query based on case sensitivity
            context_to_check = context if case_sensitive else context.lower()
            query_to_check = query if case_sensitive else query.lower()
            
            if search_query in context_to_check or search_query in query_to_check:
                results.append({
                    'url': url,
                    'context': context,
                    'query': query,
                    'original_index': i  # Store the original row index
                })
    
    elif search_type == "Regular Expression Search":
        try:
            pattern = re.compile(search_query, flags=0 if case_sensitive else re.IGNORECASE)
            for i, row in data.iterrows():
                # Make sure all values are strings and handle NaN/None values
                context = str(row['context']) if not pd.isna(row['context']) else ""
                query = str(row['query']) if not pd.isna(row['query']) else ""
                url = str(row['url']) if not pd.isna(row['url']) else ""
                
                try:
                    if pattern.search(context) or pattern.search(query):
                        results.append({
                            'url': url,
                            'context': context,
                            'query': query,
                            'original_index': i  # Store the original row index
                        })
                except (TypeError, AttributeError) as e:
                    print(f"Error searching row {i}: {e}")
                    continue
        except re.error as e:
            return f"Regular expression error: {str(e)}"
    
    # Format results
    if not results:
        return "No matching results found"
    
    # Sort results by their original index if preserve_order is True
    if preserve_order:
        results.sort(key=lambda x: x['original_index'])
    
    # Create HTML formatted results for clickable links with better styling
    formatted_results = "<div class='search-results'>"
    for i, result in enumerate(results, 1):
        url = result['url']
        url_safe = html.escape(url)
        original_idx = result['original_index'] + 1  # +1 for 1-based indexing for display
        
        formatted_results += f"<div class='result-item'>"
        formatted_results += f"<h3>Result {i} <span class='original-index'>(Dataset Row: {original_idx})</span></h3>"
        formatted_results += f"<p><b>URL:</b> <a href='{url_safe}' target='_blank'>{url_safe}</a></p>"
        formatted_results += f"<p><b>Query:</b> {html.escape(str(result['query']))}</p>"
        
        # Handle context display safely
        context = result['context']
        try:
            context_preview = context[:300] + ('...' if len(context) > 300 else '')
            context_preview = html.escape(context_preview)
        except (TypeError, AttributeError):
            context_preview = html.escape(str(context))
            
        formatted_results += f"<p><b>Context:</b> {context_preview}</p>"
        formatted_results += "</div><hr>"
    
    formatted_results += "</div>"
    return formatted_results

# Function to generate an answer using Groq API for a selected query
def generate_answer_with_groq(city, question, max_sources=3, api_key=None, temperature=0.3):
    if not api_key or api_key.strip() == "":
        return "Error: Groq API key not provided. Please enter your API key in the field above."
    
    # Try to initialize the Groq client with the provided API key
    # Handle potential proxy-related issues on Hugging Face with multiple fallback strategies
    client = None
    
    # Strategy 1: Try basic initialization
    try:
        client = Groq(api_key=api_key)
    except TypeError as e:
        if "proxies" in str(e):
            # Strategy 2: Clear proxy environment variables and try again
            try:
                import os
                proxy_vars = ['http_proxy', 'https_proxy', 'HTTP_PROXY', 'HTTPS_PROXY', 'no_proxy', 'NO_PROXY']
                original_values = {}
                
                # Store and clear all proxy-related env vars
                for var in proxy_vars:
                    if var in os.environ:
                        original_values[var] = os.environ[var]
                        del os.environ[var]
                
                # Also try to clear any requests-related proxy settings
                import sys
                if 'requests' in sys.modules:
                    import requests
                    # Clear any session-level proxy settings
                    requests.Session.proxies = {}
                
                # Try initializing again with clean environment
                client = Groq(api_key=api_key)
                
                # Restore original environment variables
                for var, value in original_values.items():
                    os.environ[var] = value
                    
            except Exception as fallback_e1:
                # Strategy 3: Try importing and using Groq differently
                try:
                    # Force reload the Groq module to clear any cached configurations
                    if 'groq' in sys.modules:
                        import importlib
                        importlib.reload(sys.modules['groq'])
                    
                    from groq import Groq as GroqClient
                    
                    # Try with explicit parameter naming
                    client = GroqClient(api_key=api_key)
                    
                except Exception as fallback_e2:
                    # Strategy 4: Try creating a minimal client configuration
                    try:
                        # Import inspect to check function signature
                        import inspect
                        groq_init_sig = inspect.signature(Groq.__init__)
                        
                        # Create kwargs with only supported parameters
                        valid_kwargs = {'api_key': api_key}
                        
                        # Only include parameters that exist in the constructor
                        for param_name in groq_init_sig.parameters:
                            if param_name in ['self']:
                                continue
                            if param_name == 'api_key':
                                valid_kwargs['api_key'] = api_key
                        
                        client = Groq(**valid_kwargs)
                        
                    except Exception as fallback_e3:
                        return (f"Error initializing Groq client after multiple attempts:\n"
                               f"Original error: {str(e)}\n"
                               f"Fallback 1 failed: {str(fallback_e1)}\n"
                               f"Fallback 2 failed: {str(fallback_e2)}\n"
                               f"Fallback 3 failed: {str(fallback_e3)}\n"
                               f"This might be due to version incompatibility or environment configuration on Hugging Face.")
        else:
            return f"Error initializing Groq client: {str(e)}"
    except Exception as e:
        return f"Error initializing Groq client: {str(e)}"
    
    # Check if client was successfully created
    if client is None:
        return "Failed to initialize Groq client after all attempts."
    
    data = all_data.get(city)
    if data is None:
        return "City data not found"
        
    # Find most relevant entries for the question
    # This is a simple relevance sorting based on TF-IDF-like scoring
    # For a production app, consider using proper embedding and semantic search
    scores = []
    
    # Keywords that indicate modern tourism/hotel content to deprioritize
    tourism_keywords = ['hotel', 'vacation', 'booking', 'resort', 'accommodation', 'travel package', 
                       'tourism', 'tourist', 'reservation', 'stay', 'room', 'suite', 'spa', 'restaurant']
    
    for i, row in data.iterrows():
        context = str(row['context']) if not pd.isna(row['context']) else ""
        url = str(row['url']) if not pd.isna(row['url']) else ""
        
        # Check if this entry is primarily about modern tourism
        context_lower = context.lower()
        url_lower = url.lower()
        tourism_score = sum(1 for keyword in tourism_keywords 
                          if keyword in context_lower or keyword in url_lower)
        
        # Simple scoring: count word overlap between question and context
        question_words = set(question.lower().split())
        context_words = set(context.lower().split())
        overlap = len(question_words.intersection(context_words))
        
        # Add a score if there are words in common, but penalize tourism content
        if overlap > 0:
            # Reduce score for entries with high tourism content
            final_score = overlap - (tourism_score * 0.5)  # Penalize tourism content
            if final_score > 0:  # Only include if still has positive relevance
                scores.append({
                    'index': i,
                    'score': final_score,
                    'url': url,
                    'context': context,
                    'tourism_score': tourism_score
                })
    
    # Sort by score (descending) and take the top entries
    scores.sort(key=lambda x: x['score'], reverse=True)
    top_entries = scores[:max_sources]
    
    if not top_entries:
        return f"No relevant information found in the {city} dataset for this question."
    
    # Build a context from the most relevant entries
    context_for_llm = f"Question about the ancient city of {city}: {question}\n\n"
    context_for_llm += "Information from dataset:\n\n"
    
    for i, entry in enumerate(top_entries, 1):
        context_for_llm += f"Source {i}: {entry['url']}\n"
        context_for_llm += f"Context: {entry['context'][:500]}...\n\n"
    
    # Create a prompt for the LLM
    prompt = f"""You are an expert historian specializing in ancient cities. 
Use the following information to answer the question about the ancient city of {city}.
Base your answer ONLY on the provided information and cite the sources.
If you cannot find relevant information to answer the question, say so honestly.

IMPORTANT: Ignore any information about modern hotels, vacation packages, tourism accommodations, travel bookings, or contemporary tourism services. Focus only on historical, archaeological, and scholarly information about the ancient city.

{context_for_llm}

Answer the question in a comprehensive, detailed, and informative way. Provide as much relevant historical context as possible. Include proper citations to the sources using [Source X] notation.
Question: {question}

First, conduct a thorough analysis of each source - evaluate the information quality, relevance, and historical significance. Skip any sources that only contain information about hotels, vacations, or modern tourism.
Then provide a detailed, well-structured answer with comprehensive explanations and proper citations focused on historical and archaeological content. Include relevant background information, context, and connections to broader historical themes when supported by the sources.

Answer with this structure:
[THINKING]
(Show your detailed analysis of the sources here, noting if any sources are skipped due to being about hotels/tourism. Explain how you're weighing the information and what historical connections you're making.)
[/THINKING]

[ANSWER]
(Your comprehensive, detailed answer with citations, focusing on historical content only. Provide thorough explanations, context, and analysis based on the available sources.)
[/ANSWER]"""

    try:
        # Make the API call to Groq
        start_time = time.time()
        
        # Call Groq API with the deepseek-r1-distill-llama-70b model
        response = client.chat.completions.create(
            model="moonshotai/kimi-k2-instruct-0905",
            messages=[
                {"role": "system", "content": "You are an expert historian specializing in ancient cities."},
                {"role": "user", "content": prompt}
            ],
            temperature=temperature,
            max_tokens=4000,  # Increased for longer, more comprehensive answers
            top_p=0.9,  
        )
        
        completion_time = time.time() - start_time
        
        # Extract and format the response
        full_response = response.choices[0].message.content
        
        # Check for explicit markers first
        thinking_match = re.search(r'\[THINKING\](.*?)\[/THINKING\]', full_response, re.DOTALL)
        answer_match = re.search(r'\[ANSWER\](.*?)\[/ANSWER\]', full_response, re.DOTALL)
        
        # Initialize variables
        thinking = ""
        answer = ""
        
        # Try different strategies to extract thinking and answer sections
        if thinking_match and answer_match:
            # Case 1: Both markers exist
            thinking = thinking_match.group(1).strip()
            answer = answer_match.group(1).strip()
        elif "Final Answer:" in full_response:
            # Case 2: There's a "Final Answer:" heading
            parts = full_response.split("Final Answer:", 1)
            thinking = parts[0].strip()
            answer = parts[1].strip()
        elif "**Analysis of Sources:**" in full_response and "**Conclusion:**" in full_response:
            # Case 3: Look for analysis section followed by conclusion
            analysis_start = full_response.find("**Analysis of Sources:**")
            conclusion_start = full_response.find("**Conclusion:**")
            if analysis_start < conclusion_start:
                thinking = full_response[:analysis_start].strip()
                answer = full_response[analysis_start:].strip()
            else:
                thinking = full_response[:conclusion_start].strip()
                answer = full_response[conclusion_start:].strip()
        elif "Thus," in full_response and "Therefore," in full_response:
            # Case 4: Look for natural language transitions
            thinking_end = max(full_response.rfind("Thus,"), full_response.rfind("Therefore,"))
            if thinking_end > 0:
                thinking = full_response[:thinking_end].strip()
                answer = full_response[thinking_end:].strip()
        elif "Starting with Source" in full_response or "Source 1" in full_response:
            # Case 5: Detect source analysis pattern
            # Look for where detailed source analysis ends and final answer begins
            patterns = [
                r"\n\n(?:To address|Based on|In conclusion|The answer|Therefore,|Thus,)",
                r"\n\n\*\*.*?\*\*",  # Look for bold headings that might start the answer
                r"\n\nGiven the",
                r"\n\nFrom the"
            ]
            
            split_point = -1
            for pattern in patterns:
                matches = list(re.finditer(pattern, full_response, re.IGNORECASE))
                if matches:
                    # Take the last match to ensure we're at the final answer section
                    split_point = matches[-1].start()
                    break
            
            if split_point > 0:
                thinking = full_response[:split_point].strip()
                answer = full_response[split_point:].strip()
            else:
                # Fallback: try to split at paragraph that doesn't start with "Source"
                parts = re.split(r'\n\n(?![Ss]ource)', full_response, 1)
                if len(parts) > 1 and len(parts[1]) > 100:  # Make sure second part is substantial
                    thinking = parts[0].strip()
                    answer = parts[1].strip()
                else:
                    thinking = "Source analysis integrated with response."
                    answer = full_response
        else:
            # Case 6: Try to split at a double newline followed by a sentence
            # that doesn't start with "Source" (which is likely part of analysis)
            parts = re.split(r'\n\n(?![Ss]ource)', full_response, 1)
            if len(parts) > 1 and len(parts[1]) > 50:  # Make sure second part is substantial
                thinking = parts[0].strip()
                answer = parts[1].strip()
            else:
                # Case 7: Default - use the whole response as answer and note no clear division
                thinking = "Analysis not clearly separated in the model's response."
                answer = full_response
        
        # Format the answer as HTML with collapsible thinking and prominent answer sections
        html_answer = f"<div class='llm-answer'>"
        
        # Add the main answer section first (most prominent)
        html_answer += "<div class='answer-section'>"
        html_answer += "<h3>Answer:</h3>"
        
        # Format answer with proper paragraphs and citation highlighting
        formatted_answer = answer
        
        # Highlight source citations [Source X]
        formatted_answer = re.sub(
            r'\[Source (\d+)\]', 
            r'<span class="citation">[Source \1]</span>', 
            formatted_answer
        )
        
        # Add paragraph breaks
        formatted_answer = formatted_answer.replace("\n\n", "</p><p>")
        formatted_answer = f"<p>{formatted_answer}</p>"
        
        html_answer += f"<div class='answer-content'>{formatted_answer}</div>"
        html_answer += "</div>"
        
        # Add the collapsible thinking section
        html_answer += "<div class='thinking-section'>"
        html_answer += """
        <details class='thinking-details'>
            <summary class='thinking-summary'>
                <span class='thinking-icon'>🔍</span>
                <span class='thinking-title'>Show Analysis Process</span>
                <span class='thinking-chevron'>▼</span>
            </summary>
            <div class='thinking-content-wrapper'>
        """
        
        # Format thinking text with proper paragraphs and source highlighting
        formatted_thinking = thinking
        
        # Replace "Source X:" with bold, highlighted version
        for i in range(1, 10):  # Support up to 9 sources
            formatted_thinking = re.sub(
                rf"Source {i}:", 
                f"<span class='source-highlight'>Source {i}:</span>", 
                formatted_thinking
            )
        
        # Add paragraph breaks for readability
        formatted_thinking = formatted_thinking.replace("\n\n", "</p><p>")
        formatted_thinking = f"<p>{formatted_thinking}</p>"
        
        html_answer += f"<div class='thinking-content'>{formatted_thinking}</div>"
        html_answer += "</div></details></div>"
        
        # Add source references at the bottom
        html_answer += "<div class='sources'><h4>Sources:</h4><ul>"
        for i, entry in enumerate(top_entries, 1):
            url_safe = html.escape(entry['url'])
            html_answer += f"<li>[Source {i}]: <a href='{url_safe}' target='_blank'>{url_safe}</a></li>"
        html_answer += "</ul></div>"
        
        # Add a small note at the bottom
        html_answer += f"<p class='model-info'><small>Generated using moonshotai/kimi-k2-instruct-0905 in {completion_time:.2f} seconds</small></p></div>"
        
        return html_answer
        
    except Exception as e:
        return f"Error generating answer: {str(e)}"

# Load all CSV files on startup
all_data = load_csv_files()
city_names = list(all_data.keys())
if not city_names:
    city_names = ["No data found"]

# Create the Gradio interface
with gr.Blocks(title="Archaeological Query Engine") as app:
       
    # Add tabs - make sure there's only one top-level Tabs component
    with gr.Tabs() as tabs:
        with gr.TabItem("Search Dataset"):
            gr.Markdown("Search through information about ancient cities from CSV files.")
            
            with gr.Row():
                with gr.Column():
                    city_dropdown = gr.Dropdown(
                        choices=city_names,
                        value=city_names[0] if city_names else None,
                        label="Select City"
                    )
                    
                    # Dropdown for queries based on the selected city
                    query_dropdown = gr.Dropdown(
                        choices=get_queries_for_city(city_names[0] if city_names else None),
                        label="Select a Query",
                        allow_custom_value=True
                    )
                    
                    search_type = gr.Radio(
                        choices=["Simple Text Search", "Regular Expression Search"],
                        value="Simple Text Search",
                        label="Search Type"
                    )
                    
                    # Keep a text box for custom queries
                    search_query = gr.Textbox(
                        label="Custom Search Query (optional)",
                        placeholder="Enter custom text to search for..."
                    )
                    
                    case_sensitive = gr.Checkbox(
                        label="Case Sensitive",
                        value=False
                    )
                    
                    show_empty_queries = gr.Checkbox(
                        label="Show Entries Without Queries",
                        value=False,
                        info="Check this to display entries that have empty or missing queries"
                    )
                    
                    preserve_order = gr.Checkbox(
                        label="Preserve Original Dataset Order",
                        value=True,
                        info="When checked, results will be displayed in their original order from the dataset. When unchecked, results will be displayed in the order they are found."
                    )
                    
                    search_button = gr.Button("Search")
                
                with gr.Column():
                    results_text = gr.HTML(
                        label="Search Results",
                        value="",
                        elem_classes=["results-output"]
                    )
                    
                    stats_text = gr.Textbox(
                        label="Dataset Statistics",
                        value=f"Total cities loaded: {len(city_names)}\nCities: {', '.join(city_names)}"
                    )
            
            # Update the query dropdown when the city changes
            def update_queries(city):
                return gr.Dropdown(choices=get_queries_for_city(city))
            
            city_dropdown.change(
                fn=update_queries,
                inputs=city_dropdown,
                outputs=query_dropdown
            )
            
            # Use either the dropdown query or the custom search query
            def search_with_queries(city, search_type, query_from_dropdown, custom_query, case_sensitive, show_empty_queries, preserve_order):
                if show_empty_queries:
                    # If show_empty_queries is checked, we show entries without queries
                    return find_empty_queries(city, preserve_order)
                else:
                    # Otherwise, use the custom query if provided, otherwise use the dropdown selection
                    final_query = custom_query if custom_query and custom_query.strip() else query_from_dropdown
                    return search_data(city, search_type, final_query, case_sensitive, preserve_order)
            
            search_button.click(
                fn=search_with_queries,
                inputs=[city_dropdown, search_type, query_dropdown, search_query, case_sensitive, show_empty_queries, preserve_order],
                outputs=results_text
            )
        
        # Add new tab for AI-generated answers using Groq API
        with gr.TabItem("AI Answers (Groq API)"):
            gr.Markdown("Ask questions about the dataset and get AI-generated answers using the Groq API with the moonshotai/kimi-k2-instruct-0905 model.")
            
            with gr.Row():
                with gr.Column():
                    # API key is now hardcoded in the code
                                     
                    ai_city_dropdown = gr.Dropdown(
                        choices=city_names,
                        value=city_names[0] if city_names else None,
                        label="Select City"
                    )
                    
                    question_input = gr.Textbox(
                        label="Ask a Question",
                        placeholder="E.g., What was the historical significance of this ancient city?",
                        lines=3
                    )
                    
                    max_sources_slider = gr.Slider(
                        minimum=1,
                        maximum=10,
                        value=3,
                        step=1,
                        label="Maximum Number of Sources to Consider",
                        info="Higher values may provide more comprehensive answers but will take longer"
                    )
                    
                    temperature_slider = gr.Slider(
                        minimum=0.0,
                        maximum=1.0,
                        value=0.3,
                        step=0.1,
                        label="Temperature",
                        info="Lower values create more focused answers, higher values create more creative ones"
                    )
                    
                    generate_button = gr.Button("Generate Answer")
                
                with gr.Column():
                    answer_output = gr.HTML(
                        label="AI-Generated Answer",
                        value="",
                        elem_classes=["results-output"]
                    )
            
        # Function to handle the Generate Answer button click
        def on_generate_answer(city, question, max_sources, temperature):
            if not question or not question.strip():
                return "Please enter a question to generate an answer."
            groq_api_key = os.environ.get("GROQ_API")
            if not groq_api_key:
                return (
                    "Error: GROQ_API environment variable not set. "
                    "Please set your Groq API key in the environment."
                )
            try:
                return generate_answer_with_groq(
                    city, question, max_sources, groq_api_key, temperature
                )
            except Exception as e:
                return f"Error: {str(e)}"

        generate_button.click(
            fn=on_generate_answer,
            inputs=[ai_city_dropdown,
                    question_input,
                    max_sources_slider,
                    temperature_slider],
            outputs=answer_output
        )

    # Add CSS styling
    gr.HTML("""
    <style>
    .gradio-container {
        font-family: 'Segoe UI', 'Arial', sans-serif;
    }
    .results-output {
        max-height: 600px;
        overflow-y: auto;
        padding: 15px;
        border: 1px solid #e2e8f0;
        border-radius: 8px;
        box-shadow: 0 2px 5px rgba(0,0,0,0.05);
    }
    a {
        color: #3182ce;
        text-decoration: none;
        transition: color 0.2s;
    }
    a:hover {
        text-decoration: underline;
        color: #2c5282;
    }
    b {
        color: #2d3748;
    }
    .search-results {
        font-family: 'Segoe UI', 'Arial', sans-serif;
    }
    .result-item {
        margin-bottom: 18px;
        padding: 15px;
        background-color: #f9f9f9;
        border-radius: 8px;
        box-shadow: 0 1px 3px rgba(0,0,0,0.05);
        transition: transform 0.2s, box-shadow 0.2s;
    }
    .result-item:hover {
        transform: translateY(-2px);
        box-shadow: 0 4px 6px rgba(0,0,0,0.1);
    }
    .original-index {
        font-size: 0.8em;
        color: #718096;
        font-weight: normal;
    }
    .result-item h3 {
        margin-top: 0;
        color: #2d3748;
        font-weight: 600;
    }
    .result-item:nth-child(odd) {
        background-color: #f5f7fa;
    }
    .result-item:nth-child(even) {
        background-color: #ffffff;
    }
    hr {
        border: 0;
        height: 1px;
        background-color: #e2e8f0;
        margin: 20px 0;
    }
    .llm-answer {
        font-family: 'Segoe UI', 'Arial', sans-serif;
        line-height: 1.7;
        padding: 20px;
        border-radius: 12px;
        background-color: #f8fafc;
        box-shadow: 0 4px 12px rgba(0,0,0,0.08);
        border: 1px solid #cbd5e0;
        margin: 10px 0;
    }
    .answer-section {
        margin-bottom: 25px;
        background-color: #ffffff;
        padding: 25px;
        border-radius: 10px;
        border: 1px solid #e2e8f0;
        box-shadow: 0 2px 8px rgba(0,0,0,0.04);
    }
    .answer-section h3 {
        margin-top: 0;
        color: #1a202c;
        font-weight: 700;
        font-size: 1.3em;
        margin-bottom: 20px;
        display: flex;
        align-items: center;
    }
    .answer-section h3::before {
        content: "💡";
        margin-right: 10px;
        font-size: 1.1em;
    }
    .answer-content {
        font-size: 1.05em;
        line-height: 1.8;
        color: #2d3748;
    }
    .answer-content p {
        margin-bottom: 16px;
    }
    .answer-content .citation {
        font-weight: 600;
        color: #3182ce;
        background-color: #ebf8ff;
        padding: 2px 8px;
        border-radius: 6px;
        font-size: 0.9em;
        border: 1px solid #bee3f8;
    }
    .thinking-section {
        margin-bottom: 20px;
    }
    .thinking-details {
        background-color: #f7fafc;
        border: 1px solid #e2e8f0;
        border-radius: 8px;
        overflow: hidden;
    }
    .thinking-summary {
        cursor: pointer;
        padding: 15px 20px;
        background-color: #edf2f7;
        border-bottom: 1px solid #e2e8f0;
        display: flex;
        align-items: center;
        font-weight: 600;
        color: #4a5568;
        transition: background-color 0.2s ease;
        user-select: none;
    }
    .thinking-summary:hover {
        background-color: #e2e8f0;
    }
    .thinking-icon {
        margin-right: 10px;
        font-size: 1.1em;
    }
    .thinking-title {
        flex-grow: 1;
        font-size: 0.95em;
    }
    .thinking-chevron {
        font-size: 0.8em;
        transition: transform 0.3s ease;
        margin-left: 10px;
    }
    .thinking-details[open] .thinking-chevron {
        transform: rotate(180deg);
    }
    .thinking-content-wrapper {
        padding: 0;
    }
    .thinking-content {
        background-color: #f0f4f8;
        padding: 20px;
        margin: 0;
        font-size: 0.93em;
        line-height: 1.6;
        color: #4a5568;
    }
    .thinking-content p {
        margin-bottom: 12px;
    }
    .thinking-content .source-highlight {
        font-weight: 600;
        color: #2b6cb0;
        background-color: #ebf4ff;
        padding: 2px 6px;
        border-radius: 4px;
        border: 1px solid #bee3f8;
    }
    .sources {
        font-size: 0.95em;
        margin-top: 25px;
        padding: 20px;
        background-color: #ffffff;
        border-radius: 8px;
        border: 1px solid #e2e8f0;
        color: #4a5568;
    }
    .sources h4 {
        margin-top: 0;
        color: #2d3748;
        font-weight: 600;
        font-size: 1.05em;
        margin-bottom: 15px;
        display: flex;
        align-items: center;
    }
    .sources h4::before {
        content: "📚";
        margin-right: 8px;
        font-size: 1em;
    }
    .sources ul {
        padding-left: 20px;
        margin: 0;
    }
    .sources li {
        margin-bottom: 8px;
        line-height: 1.5;
    }
    .model-info {
        text-align: right;
        color: #718096;
        margin-top: 20px;
        margin-bottom: 0;
        font-size: 0.85em;
        padding-top: 15px;
        border-top: 1px solid #e2e8f0;
    }
    </style>
    """)

# Launch the app
if __name__ == "__main__":
    try:
        print(f"Loaded {len(city_names)} cities: {', '.join(city_names)}")
        app.launch(show_error=True)
    except Exception as e:
        print(f"Error starting application: {e}")
        import traceback
        traceback.print_exc()