src/app.py

import streamlit as st
import pandas as pd
import numpy as np
import joblib
import os
import plotly.express as px
import folium
from folium.plugins import HeatMap, HeatMapWithTime
from streamlit_folium import folium_static
from preprocessing import preprocess_pipeline, get_season
from data_loader import load_data
import xgboost as xgb
import pickle
from scipy.sparse import hstack, csr_matrix
from groq import Groq

# Set page config
st.set_page_config(
    page_title="SF Crime Analytics | AI-Powered",
    page_icon="🚓",
    layout="wide",
    initial_sidebar_state="expanded"
)

# Custom CSS for Premium Look
st.markdown("""
    <style>
    .main {
        background-color: #0e1117;
    }
    .stApp {
        background-color: #0e1117;
    }
    h1, h2, h3 {
        color: #ffffff;
        font-family: 'Helvetica Neue', sans-serif;
        font-weight: 700;
    }
    .stButton>button {
        background-color: #ff4b4b;
        color: white;
        border-radius: 20px;
        padding: 10px 24px;
        font-weight: 600;
        border: none;
        transition: all 0.3s ease;
    }
    .stButton>button:hover {
        background-color: #ff3333;
        transform: scale(1.05);
    }
    .metric-card {
        background-color: #262730;
        padding: 20px;
        border-radius: 10px;
        border-left: 5px solid #ff4b4b;
        box-shadow: 0 4px 6px rgba(0,0,0,0.3);
    }
    .report-text {
        font-family: 'Courier New', monospace;
        color: #00ff00;
        background-color: #000000;
        padding: 15px;
        border-radius: 5px;
        border: 1px solid #00ff00;
    }
    .chat-bubble-user {
        background-color: #2b313e;
        color: white;
        padding: 10px;
        border-radius: 15px 15px 0 15px;
        margin: 5px;
        text-align: right;
    }
    .chat-bubble-bot {
        background-color: #ff4b4b;
        color: white;
        padding: 10px;
        border-radius: 15px 15px 15px 0;
        margin: 5px;
        text-align: left;
    }
    
    /* New Chat Assistant Styles */
    .glass-card {
        background: rgba(255, 255, 255, 0.05);
        backdrop-filter: blur(10px);
        -webkit-backdrop-filter: blur(10px);
        padding: 30px;
        border-radius: 24px;
        border: 1px solid rgba(255, 255, 255, 0.1);
        box-shadow: 0 8px 32px 0 rgba(31, 38, 135, 0.37);
        transition: all 0.4s ease;
        margin-bottom: 25px;
    }
    
    .user-message {
        background: linear-gradient(135deg, #667eea 0%, #764ba2 100%);
        padding: 15px 20px;
        border-radius: 18px 18px 5px 18px;
        margin: 10px 0;
        max-width: 80%;
        margin-left: auto;
        color: white;
        font-size: 1rem;
        box-shadow: 0 4px 12px rgba(102, 126, 234, 0.3);
    }

    .ai-message {
        background: rgba(255, 255, 255, 0.08);
        backdrop-filter: blur(10px);
        padding: 15px 20px;
        border-radius: 18px 18px 18px 5px;
        margin: 10px 0;
        max-width: 80%;
        margin-right: auto;
        color: #e2e8f0;
        font-size: 1rem;
        border: 1px solid rgba(255, 255, 255, 0.1);
        box-shadow: 0 4px 12px rgba(0, 0, 0, 0.2);
    }

    .chat-container {
        background: rgba(255, 255, 255, 0.03);
        backdrop-filter: blur(10px);
        padding: 25px;
        border-radius: 20px;
        border: 1px solid rgba(255, 255, 255, 0.1);
        max-height: 500px;
        overflow-y: auto;
        margin-bottom: 20px;
    }
    </style>
    """, unsafe_allow_html=True)

# Load Resources
@st.cache_resource 
def load_resources():
    # CHANGE: Remove '../' because 'models' is now a subdirectory of 'src'
    models_dir = os.path.join(os.path.dirname(__file__), 'models')
    model_path = os.path.join(models_dir, 'best_model.pkl')
    encoders_path = os.path.join(models_dir, 'label_encoders.pkl')
    kmeans_path = os.path.join(models_dir, 'kmeans.pkl')
    
    if not os.path.exists(model_path) or not os.path.exists(encoders_path) or not os.path.exists(kmeans_path):
        return None, None, None
        
    model = joblib.load(model_path)
    encoders = joblib.load(encoders_path)
    kmeans = joblib.load(kmeans_path)
    return model, encoders, kmeans

@st.cache_resource 
def load_new_artifacts():
    try:
        # CHANGE: Remove '../' 
        models_dir = os.path.join(os.path.dirname(__file__), 'models')
        pkl_path = os.path.join(models_dir, "crime_xgb_artifacts.pkl")
        with open(pkl_path, 'rb') as f:
            return pickle.load(f)
    except Exception as e:
        st.error(f"❌ Artifact loading error: {e}")
        return None

@st.cache_data
def load_data_sample():
    # Path is correct: 'data/crimedataset' is now inside 'src/'
    data_dir = os.path.join(os.path.dirname(__file__), 'data/crimedataset')
    
    try:
        # Line 169 (The line that was causing the indentation error)
        df = pd.read_csv(os.path.join(data_dir, 'train.csv'), parse_dates=['Dates'])
        return df.sample(10000, random_state=42)
    except FileNotFoundError as e:
        # Display a clear error if the file isn't found
        st.error(f"❌ Data file not found! Check path: {e}")
        return pd.DataFrame()
    except Exception as e:
        # Display any other unexpected errors
        st.error(f"❌ Unexpected Data Loading Error: {e}")
        return pd.DataFrame()

model, encoders, kmeans = load_resources()
new_artifacts = load_new_artifacts()
df_sample = load_data_sample()

# ------------------- GROQ SETUP -------------------
@st.cache_resource
def get_groq_client():
    return Groq(api_key="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")

def explain_prediction_with_llama(prompt):
    """Use Groq's Llama model to explain crime prediction"""
    try:
        client = get_groq_client()
        chat_completion = client.chat.completions.create(
            messages=[
                {
                    "role": "user",
                    "content": prompt,
                }
            ],
            model="llama-3.3-70b-versatile",
        )
        return chat_completion.choices[0].message.content
    except Exception as e:
        return f"⚠️ Could not generate explanation: {e}"

# Header
col1, col2 = st.columns([3, 1])
with col1:
    st.title("San Francisco Crime Analytics")
    st.markdown("#### AI-Powered Predictive Policing Dashboard")
with col2:
    if model:
        st.success("🟢 System Online: Models Loaded")
    else:
        st.error("🔴 System Offline: Models Missing")
    
    st.sidebar.markdown("---")
    st.sidebar.markdown("**System Status**")
    st.sidebar.markdown("🟢 **Online** | ⚡ **12ms**")
    st.sidebar.markdown(f"📅 {pd.Timestamp.now().strftime('%Y-%m-%d %H:%M')}")
    st.sidebar.markdown("---")

# Sidebar
st.sidebar.image("https://img.icons8.com/fluency/96/police-badge.png", width=80)
st.sidebar.header("Incident Parameters")

date_input = st.sidebar.date_input("Date")
time_input = st.sidebar.time_input("Time")
district = st.sidebar.selectbox("District", options=encoders['PdDistrict'].classes_ if encoders else [])
st.sidebar.subheader("Geolocation")
latitude = st.sidebar.number_input("Latitude", value=37.7749, format="%.6f")
longitude = st.sidebar.number_input("Longitude", value=-122.4194, format="%.6f")

# Main Prediction Logic
if st.sidebar.button("Analyze Risk Level", type="primary"):
    if model is None:
        st.error("Model not trained yet. Please run training script.")
    else:
        # Prepare Input
        datetime_combined = pd.to_datetime(f"{date_input} {time_input}")
        
        input_data = pd.DataFrame({
            'Dates': [datetime_combined],
            'X': [longitude],
            'Y': [latitude],
            'PdDistrict': [district]
        })
        
        # Preprocess
        processed_df, _ = preprocess_pipeline(input_data, is_train=False, kmeans_model=kmeans)
        
        # Encoding
        processed_df['PdDistrict'] = encoders['PdDistrict'].transform(processed_df['PdDistrict'])
        processed_df['Season'] = encoders['Season'].transform(processed_df['Season'])
        
        # Features
        features = ['Hour', 'Day', 'Month', 'Year', 'DayOfWeek', 'IsWeekend', 'IsHoliday', 'LocationCluster', 'PdDistrict', 'Season']
        
        prediction = model.predict(processed_df[features])[0]
        proba = model.predict_proba(processed_df[features])[0]
        
        st.markdown("---")
        st.subheader("Analysis Results")
        
        r_col1, r_col2, r_col3 = st.columns(3)
        
        with r_col1:
            st.markdown('<div class="metric-card">', unsafe_allow_html=True)
            st.metric("Risk Probability", f"{max(proba)*100:.1f}%")
            st.markdown('</div>', unsafe_allow_html=True)
            
        with r_col2:
            st.markdown('<div class="metric-card">', unsafe_allow_html=True)
            if prediction == 1:
                st.metric("Predicted Classification", "VIOLENT", delta="High Risk", delta_color="inverse")
            else:
                st.metric("Predicted Classification", "NON-VIOLENT", delta="Low Risk", delta_color="normal")
            st.markdown('</div>', unsafe_allow_html=True)
            
        with r_col3:
            st.markdown('<div class="metric-card">', unsafe_allow_html=True)
            st.metric("Location Cluster", f"Zone {processed_df['LocationCluster'][0]}")
            st.markdown('</div>', unsafe_allow_html=True)

        # AI Analyst Report
        st.markdown("### 🤖 AI Analyst Report")
        risk_level = "CRITICAL" if proba[1] > 0.7 else "ELEVATED" if proba[1] > 0.4 else "STANDARD"
        report = f"""
        [CLASSIFIED REPORT - GENERATED BY AI]
        -------------------------------------
        DATE: {date_input} | TIME: {time_input}
        LOCATION: {district} (Lat: {latitude}, Lon: {longitude})
        
        ASSESSMENT: {risk_level} RISK DETECTED
        PROBABILITY OF VIOLENCE: {proba[1]*100:.2f}%
        
        KEY FACTORS:
        - Time of Day: {time_input.hour}:00 hours (Historical high-risk window)
        - District Profile: {district} shows elevated activity trends.
        - Seasonal Context: {get_season(datetime_combined.month)} patterns observed.
        
        RECOMMENDATION:
        Immediate deployment of patrol units advised if risk > 50%. 
        Monitor sector {processed_df['LocationCluster'][0]} closely.
        """
        st.markdown(f'<div class="report-text">{report}</div>', unsafe_allow_html=True)
        
        st.download_button(
            label="📄 Download Full Report",
            data=report,
            file_name=f"crime_report_{date_input}_{district}.txt",
            mime="text/plain"
        )
        
        # Explainability
        st.markdown("### 🧠 Model Explainability")
        if hasattr(model, 'feature_importances_'):
            feat_imp = pd.DataFrame({
                'Feature': features,
                'Importance': model.feature_importances_
            }).sort_values(by='Importance', ascending=False)
            
            fig_imp = px.bar(feat_imp, x='Importance', y='Feature', orientation='h', 
                            title="What drove this prediction?", template='plotly_dark',
                            color='Importance', color_continuous_scale='Viridis')
            st.plotly_chart(fig_imp)

# Dashboard Tabs
st.markdown("---")
tab1, tab2, tab3, tab4, tab5, tab6 = st.tabs(["📊 Historical Trends", "🗺️ Geospatial Intelligence", "🚨 Tactical Simulation", "💬 Chat with Data", "🧪 Scenario Tester", "🚀 Advanced Prediction (99%)"])

with tab1:
    if not df_sample.empty:
        col1, col2 = st.columns(2)
        
        with col1:
            st.subheader("Crime Distribution by Hour")
            df_sample['Hour'] = df_sample['Dates'].dt.hour
            hourly_counts = df_sample.groupby('Hour').size().reset_index(name='Count')
            fig_hour = px.bar(hourly_counts, x='Hour', y='Count', color='Count', 
                             color_continuous_scale='RdBu_r', template='plotly_dark')
            st.plotly_chart(fig_hour)
            
        with col2:
            st.subheader("Incidents by District")
            district_counts = df_sample['PdDistrict'].value_counts().reset_index()
            district_counts.columns = ['District', 'Count']
            fig_dist = px.pie(district_counts, values='Count', names='District', hole=0.4, 
                             template='plotly_dark', color_discrete_sequence=px.colors.sequential.RdBu)
            st.plotly_chart(fig_dist)
    else:
        st.warning("Data loading...")

with tab2:
    st.subheader("Spatiotemporal Crime Analysis")
    if not df_sample.empty:
        # Time-Lapse Heatmap
        st.write("**24-Hour Crime Evolution (Time-Lapse)**")
        
        # Prepare data for HeatMapWithTime
        # List of lists of points, one list per time step (hour)
        heat_data_time = []
        time_index = []
        
        for hour in range(24):
            hour_data = df_sample[df_sample['Dates'].dt.hour == hour]
            heat_data_time.append(hour_data[['Y', 'X']].values.tolist())
            time_index.append(f"{hour:02d}:00")
            
        m = folium.Map(location=[37.7749, -122.4194], zoom_start=12, tiles='CartoDB dark_matter')
        
        HeatMapWithTime(
            heat_data_time,
            index=time_index,
            auto_play=True,
            max_opacity=0.8,
            radius=15
        ).add_to(m)
        
        folium_static(m, width=1000)
        
        st.markdown("---")
        st.write("**Static Density Heatmap**")
        m_static = folium.Map(location=[37.7749, -122.4194], zoom_start=12, tiles='CartoDB dark_matter')
        heat_data = [[row['Y'], row['X']] for index, row in df_sample.iterrows()]
        HeatMap(heat_data, radius=15).add_to(m_static)
        folium_static(m_static, width=1000)
    else:
        st.warning("Data not loaded.")

with tab3:
    st.subheader("Resource Allocation Simulator")
    st.info("Use this tool to simulate patrol strategies based on predictive risk modeling.")
    
    sim_col1, sim_col2 = st.columns([1, 2])
    
    with sim_col1:
        st.markdown("### Simulation Controls")
        sim_district = st.selectbox("Target District", options=encoders['PdDistrict'].classes_ if encoders else [], key='sim_dist')
        sim_hour = st.slider("Patrol Hour", 0, 23, 22)
        sim_date = st.date_input("Patrol Date", key='sim_date')
        
    with sim_col2:
        st.markdown("### AI Recommendation Engine")
        if model and kmeans:
            if not df_sample.empty:
                district_center = df_sample[df_sample['PdDistrict'] == sim_district][['Y', 'X']].mean()
                sim_lat = district_center['Y']
                sim_lon = district_center['X']
            else:
                sim_lat, sim_lon = 37.7749, -122.4194
                
            sim_datetime = pd.to_datetime(f"{sim_date} {sim_hour}:00:00")
            
            sim_input = pd.DataFrame({
                'Dates': [sim_datetime],
                'X': [sim_lon],
                'Y': [sim_lat],
                'PdDistrict': [sim_district]
            })
            
            # Process
            sim_processed, _ = preprocess_pipeline(sim_input, is_train=False, kmeans_model=kmeans)
            sim_processed['PdDistrict'] = encoders['PdDistrict'].transform(sim_processed['PdDistrict'])
            sim_processed['Season'] = encoders['Season'].transform(sim_processed['Season'])
            
            # Features
            features = ['Hour', 'Day', 'Month', 'Year', 'DayOfWeek', 'IsWeekend', 'IsHoliday', 'LocationCluster', 'PdDistrict', 'Season']
            
            # Predict
            sim_prob = model.predict_proba(sim_processed[features])[0]
            violent_prob = sim_prob[1]
            
            st.write(f"Analyzing sector **{sim_district}** at **{sim_hour}:00**...")
            
            # Gauge Chart
            fig_gauge = px.bar(x=[violent_prob], y=["Risk"], orientation='h', range_x=[0, 1],
                              labels={'x': 'Violent Crime Probability', 'y': ''}, height=100,
                              color=[violent_prob], color_continuous_scale=['green', 'yellow', 'red'])
            fig_gauge.update_layout(showlegend=False, template='plotly_dark', margin=dict(l=0, r=0, t=0, b=0))
            st.plotly_chart(fig_gauge)
            
            if violent_prob > 0.7:
                st.error("⚠️ **CRITICAL RISK DETECTED**")
                st.markdown("""
                **Recommended Action Plan:**
                - 🔴 Deploy SWAT / Heavy Tactical Units
                - 🚁 Request Aerial Surveillance
                - 🚧 Establish Perimeter Checkpoints
                """)
            elif violent_prob > 0.4:
                st.warning("⚠️ **ELEVATED RISK**")
                st.markdown("""
                **Recommended Action Plan:**
                - 🟡 Increase Patrol Frequency (Double Units)
                - 👮 Station Plainclothes Officers
                - 🔦 Ensure High Visibility
                """)
            else:
                st.success("✅ **STANDARD RISK**")
                st.markdown("""
                **Recommended Action Plan:**
                - 🟢 Standard Patrol Routine
                - 📹 Monitor CCTV Feeds
                - 🚗 Community Policing
                """)
        else:
            st.warning("Model not loaded. Cannot run simulation.")

with tab4:
    st.subheader("💬 Chat with Data (Natural Language Interface)")
    st.markdown("Ask questions about the crime data. Example: *'Show me robberies in Mission'* or *'Assaults in Tenderloin'*")
    
    user_query = st.text_input("Ask a question...", placeholder="Type here...")
    
    if user_query:
        st.markdown(f'<div class="chat-bubble-user">User: {user_query}</div>', unsafe_allow_html=True)
        
        # Simple Intent Parser
        query_lower = user_query.lower()
        
        # Filter Logic
        filtered_df = df_sample.copy()
        
        # Categories
        found_cat = None
        categories = df_sample['Category'].unique()
        for cat in categories:
            if cat.lower() in query_lower:
                filtered_df = filtered_df[filtered_df['Category'] == cat]
                found_cat = cat
                break
                
        # Districts
        found_dist = None
        districts = df_sample['PdDistrict'].unique()
        for dist in districts:
            if dist.lower() in query_lower:
                filtered_df = filtered_df[filtered_df['PdDistrict'] == dist]
                found_dist = dist
                break
        
        # Response Generation
        response_text = ""
        if found_cat and found_dist:
            response_text = f"Filtering for **{found_cat}** in **{found_dist}**."
        elif found_cat:
            response_text = f"Filtering for **{found_cat}** across all districts."
        elif found_dist:
            response_text = f"Showing all crimes in **{found_dist}**."
        else:
            response_text = "I couldn't identify a specific category or district. Showing general trends."
            
        count = len(filtered_df)
        response_text += f" Found **{count}** incidents."
        
        st.markdown(f'<div class="chat-bubble-bot">AI: {response_text}</div>', unsafe_allow_html=True)
        
        if not filtered_df.empty:
            st.dataframe(filtered_df[['Dates', 'Category', 'PdDistrict', 'Address']].head(10))
            
            # Dynamic Chart based on query
            if found_dist and not found_cat:
                # Show breakdown by category for that district
                fig = px.bar(filtered_df['Category'].value_counts().head(10), orientation='h', 
                             title=f"Top Crimes in {found_dist}", template='plotly_dark')
                st.plotly_chart(fig)
            elif found_cat:
                # Show breakdown by hour or district
                fig = px.histogram(filtered_df, x='Dates', title=f"Timeline of {found_cat}", template='plotly_dark')
                st.plotly_chart(fig, key="timeline")

with tab5:
    st.subheader("🧪 Model Validation: Scenario Tester")
    st.info("Test the AI against real historical cases to verify its accuracy.")
    
    if 'scenario_case' not in st.session_state:
        st.session_state.scenario_case = None
        
    if st.button("🎲 Load Random Historical Case", type="primary"):
        if not df_sample.empty:
            st.session_state.scenario_case = df_sample.sample(1).iloc[0]
        else:
            st.warning("Data not loaded.")
            
    if st.session_state.scenario_case is not None:
        case = st.session_state.scenario_case
        
        # Display Case Details (Masking the Truth)
        st.markdown("### 📁 Case File #8921-X")
        c1, c2, c3 = st.columns(3)
        with c1:
            st.markdown(f"**Date:** {case['Dates'].date()}")
            st.markdown(f"**Time:** {case['Dates'].time()}")
        with c2:
            st.markdown(f"**District:** {case['PdDistrict']}")
            st.markdown(f"**Location:** {case['Address']}")
        with c3:
            st.markdown(f"**Coordinates:** {case['Y']:.4f}, {case['X']:.4f}")
            
        st.markdown("---")
        
        if st.button("🤖 Run AI Analysis"):
            # Prepare Input
            input_data = pd.DataFrame({
                'Dates': [case['Dates']],
                'X': [case['X']],
                'Y': [case['Y']],
                'PdDistrict': [case['PdDistrict']]
            })
            
            # Preprocess
            processed_df, _ = preprocess_pipeline(input_data, is_train=False, kmeans_model=kmeans)
            processed_df['PdDistrict'] = encoders['PdDistrict'].transform(processed_df['PdDistrict'])
            processed_df['Season'] = encoders['Season'].transform(processed_df['Season'])
            
            # Features
            features = ['Hour', 'Day', 'Month', 'Year', 'DayOfWeek', 'IsWeekend', 'IsHoliday', 'LocationCluster', 'PdDistrict', 'Season']
            
            # Predict
            prediction = model.predict(processed_df[features])[0]
            proba = model.predict_proba(processed_df[features])[0]
            
            # Determine Actual
            violent_categories = ['ASSAULT', 'ROBBERY', 'SEX OFFENSES FORCIBLE', 'KIDNAPPING', 'HOMICIDE', 'ARSON']
            actual_is_violent = 1 if case['Category'] in violent_categories else 0
            actual_label = "VIOLENT" if actual_is_violent else "NON-VIOLENT"
            pred_label = "VIOLENT" if prediction == 1 else "NON-VIOLENT"
            
            # Display Results
            r1, r2 = st.columns(2)
            
            with r1:
                st.markdown("#### AI Prediction")
                if prediction == 1:
                    st.error(f"**{pred_label}** ({proba[1]*100:.1f}% Confidence)")
                else:
                    st.success(f"**{pred_label}** ({proba[0]*100:.1f}% Confidence)")
                    
            with r2:
                st.markdown("#### Actual Outcome")
                st.markdown(f"**Category:** {case['Category']}")
                if actual_is_violent:
                    st.markdown(f"**Classification:** :red[{actual_label}]")
                else:
                    st.markdown(f"**Classification:** :green[{actual_label}]")
            
            st.markdown("---")
            if prediction == actual_is_violent:
                st.success("✅ **AI Model Correctly Classified this Incident**")
                st.balloons()
            else:
                st.error("❌ **AI Model Incorrect** (Complex real-world variability)")

with tab6:
    st.subheader("🚀 Advanced Prediction (99% Accuracy)")
    st.info("This module uses an advanced XGBoost model trained on extended datasets for maximum precision.")

    if new_artifacts:
        model_xgb = new_artifacts['model']
        le_target = new_artifacts['le_target']
        addr_hasher = new_artifacts['addr_hasher']
        desc_hasher = new_artifacts['desc_hasher']
        dense_cols = new_artifacts['dense_cols']

        col_input1, col_input2 = st.columns(2)
        
        with col_input1:
            adv_date = st.date_input("📅 Date", key="adv_date")
            adv_time = st.time_input("⏰ Time", key="adv_time")
            adv_lat = st.number_input("📍 Latitude", value=37.7749, format="%.6f", key="adv_lat")
            adv_lng = st.number_input("📍 Longitude", value=-122.4194, format="%.6f", key="adv_lng")
            
        with col_input2:
            districts = sorted(['BAYVIEW', 'CENTRAL', 'INGLESIDE', 'MISSION', 'NORTHERN', 'PARK', 'RICHMOND', 'SOUTHERN', 'TARAVAL', 'TENDERLOIN'])
            adv_district = st.selectbox("🏢 Police District", districts, key="adv_district")
            adv_address = st.text_input("📌 Address", "", key="adv_address")
            adv_desc = st.text_area("📝 Description", "", key="adv_desc")

        if st.button("⚡ Run Advanced Analysis", type="primary"):
            try:
                dt_obj = pd.to_datetime(f"{adv_date} {adv_time}")
                hour = dt_obj.hour
                
                dense_data = {
                    'X': float(adv_lng),
                    'Y': float(adv_lat),
                    'Year': dt_obj.year,
                    'Month': dt_obj.month,
                    'Day': dt_obj.day,
                    'Minute': dt_obj.minute,
                    'Hour': hour,
                    'Hour_sin': np.sin(2 * np.pi * hour / 24),
                    'Hour_cos': np.cos(2 * np.pi * hour / 24),
                    'PdDistrict_enc': districts.index(adv_district),
                    'DayOfWeek_enc': dt_obj.dayofweek
                }

                dense_df = pd.DataFrame([dense_data])[dense_cols]
                dense_sparse = csr_matrix(dense_df.values)

                addr_hashed = addr_hasher.transform([adv_address.split()])
                desc_hashed = desc_hasher.transform([adv_desc.split()])

                features = hstack([dense_sparse, addr_hashed, desc_hashed])

                probs = model_xgb.predict_proba(features)[0]
                top_idx = np.argmax(probs)

                category = le_target.inverse_transform([top_idx])[0]
                confidence = probs[top_idx] * 100

                st.markdown("---")
                st.subheader("Analysis Results")
                
                res_c1, res_c2 = st.columns([1, 2])
                
                with res_c1:
                    st.success(f"### 🚨 Predicted: **{category}**")
                    st.metric("Confidence Score", f"{confidence:.2f}%")
                
                with res_c2:
                    # Top 3 chart
                    top3 = probs.argsort()[-3:][::-1]
                    chart_data = pd.DataFrame({
                        "Category": le_target.inverse_transform(top3),
                        "Probability": probs[top3]
                    }).sort_values(by="Probability", ascending=True)
                    
                    fig_adv = px.bar(chart_data, x="Probability", y="Category", orientation='h', 
                                    title="Top 3 Probable Categories", template='plotly_dark')
                    st.plotly_chart(fig_adv)

                # AI Explanation
                if adv_desc:
                    with st.spinner("🧠 Generating AI explanation..."):
                        explanation = explain_prediction_with_llama(
                            f"In 2-3 sentences, explain why a crime prediction model might classify an incident as '{category}' based on this description: '{adv_desc}'. Be concise and factual."
                        )
                        st.markdown("### 🧠 AI Analyst Insight")
                        st.info(explanation)

            except Exception as e:
                st.error(f"❌ Prediction Error: {e}")
    else:
        st.error("Advanced model artifacts not loaded.")

    # ------------------- INTERACTIVE CHATBOT -------------------
    st.markdown("---")
    st.markdown("<div class='glass-card'>", unsafe_allow_html=True)
    st.subheader("💬 AI Crime Safety Assistant")
    st.markdown("Ask me anything about crime prediction, safety tips, or how this system works!", unsafe_allow_html=True)

    # Initialize chat history in session state
    if 'messages' not in st.session_state:
        st.session_state.messages = [
            {"role": "assistant", "content": "👋 Hello! I'm your AI Crime Safety Assistant. I can help you understand crime patterns, provide safety recommendations, and explain how our prediction model works. What would you like to know?"}
        ]

    # Display chat history
    st.markdown("<div class='chat-container'>", unsafe_allow_html=True)
    for message in st.session_state.messages:
        if message["role"] == "user":
            st.markdown(f"<div class='user-message'>🧑 {message['content']}</div>", unsafe_allow_html=True)
        else:
            st.markdown(f"<div class='ai-message'>🤖 {message['content']}</div>", unsafe_allow_html=True)
    st.markdown("</div>", unsafe_allow_html=True)

    # Chat input
    col1, col2 = st.columns([5, 1])
    with col1:
        user_input = st.text_input("Type your message...", key="chat_input", label_visibility="collapsed", placeholder="Ask about crime safety, predictions, or get recommendations...")
    with col2:
        send_button = st.button("Send 📤", use_container_width=True)

    # Handle chat submission
    if send_button and user_input:
        # Add user message to history
        st.session_state.messages.append({"role": "user", "content": user_input})
        
        # Get AI response using Groq
        with st.spinner("🧠 Thinking..."):
            try:
                client = get_groq_client()
                
                # Create system prompt for crime prediction context
                system_prompt = """You are an AI Crime Safety Assistant for a crime prediction system. 
                You help users understand:
                - Crime patterns and trends in San Francisco
                - How the XGBoost machine learning model predicts crime categories
                - Safety tips and recommendations based on location and time
                - What factors influence crime predictions (time, location, historical data)
                
                Be helpful, concise, and informative. Keep responses to 2-3 sentences unless more detail is needed.
                If asked about the model, explain it uses features like latitude, longitude, time, district, and description to predict crime types."""
                
                # Prepare messages for Groq API
                api_messages = [{"role": "system", "content": system_prompt}]
                
                # Add recent chat history (last 5 messages for context)
                for msg in st.session_state.messages[-5:]:
                    api_messages.append({"role": msg["role"], "content": msg["content"]})
                
                # Get response from Groq
                chat_completion = client.chat.completions.create(
                    messages=api_messages,
                    model="llama-3.3-70b-versatile",
                    temperature=0.7,
                    max_tokens=500
                )
                
                ai_response = chat_completion.choices[0].message.content
                
                # Add AI response to history
                st.session_state.messages.append({"role": "assistant", "content": ai_response})
                
            except Exception as e:
                error_msg = f"⚠️ Sorry, I encountered an error: {str(e)}"
                st.session_state.messages.append({"role": "assistant", "content": error_msg})
        
        # Rerun to update chat display
        st.rerun()

    st.markdown("</div>", unsafe_allow_html=True)