app_meta_analysis.py

import streamlit as st
import pandas as pd
import numpy as np
import uuid
from datetime import datetime
import os
import io

# 頁面配置
st.set_page_config(
    page_title="Bayesian Meta-Analysis - Pokémon Type Comparison",
    page_icon="🎲",
    layout="wide",
    initial_sidebar_state="expanded"
)

# 自定義 CSS
st.markdown("""
<style>
    .streamlit-expanderHeader {
        background-color: #e8f1f8;
        border: 1px solid #b0cfe8;
        border-radius: 5px;
        font-weight: 600;
        color: #1b4f72;
    }
    .streamlit-expanderHeader:hover {
        background-color: #d0e7f8;
    }
    .stMetric {
        background-color: #f8fbff;
        padding: 10px;
        border-radius: 5px;
        border: 1px solid #d0e4f5;
    }
    .stButton > button {
        width: 100%;
        border-radius: 20px;
        font-weight: 600;
        transition: all 0.3s ease;
    }
    .stButton > button:hover {
        transform: translateY(-2px);
        box-shadow: 0 4px 8px rgba(0,0,0,0.2);
    }
    .success-box {
        background-color: #d4edda;
        border: 1px solid #c3e6cb;
        border-radius: 5px;
        padding: 10px;
        margin: 10px 0;
    }
    .info-box {
        background-color: #d1ecf1;
        border: 1px solid #bee5eb;
        border-radius: 5px;
        padding: 10px;
        margin: 10px 0;
    }
</style>
""", unsafe_allow_html=True)

# 導入自定義模組
from meta_analysis_core import BayesianMetaAnalyzer
from meta_analysis_llm_assistant import MetaAnalysisLLMAssistant
from meta_analysis_utils import (
    plot_trace_combined,
    plot_posterior,
    plot_forest,
    plot_dag,
    create_dag_legend_table,
    format_summary_stats
)

# 初始化 session state
if 'session_id' not in st.session_state:
    st.session_state.session_id = str(uuid.uuid4())
if 'analysis_results' not in st.session_state:
    st.session_state.analysis_results = None
if 'analyzer' not in st.session_state:
    st.session_state.analyzer = None
if 'available_types' not in st.session_state:
    st.session_state.available_types = []
if 'df_full' not in st.session_state:
    st.session_state.df_full = None
if 'chat_history' not in st.session_state:
    st.session_state.chat_history = []
if 'llm_assistant' not in st.session_state:
    st.session_state.llm_assistant = None

# 標題
st.title("🎲 Bayesian Meta-Analysis System")
st.markdown("### 貝氏統合分析模型")
st.markdown("---")

# Sidebar
with st.sidebar:
    st.header("⚙️ 配置設定")
    
    # API 選擇
    st.subheader("🤖 AI 助手設定")
    api_choice = st.radio(
        "選擇 LLM API",
        options=["Google Gemini", "Anthropic Claude"],
        index=0,
        help="選擇要使用的 AI 助手"
    )
    
    # API Key 輸入
    if api_choice == "Google Gemini":
        api_key = st.text_input(
            "Google Gemini API Key",
            type="password",
            help="輸入您的 Google Gemini API Key"
        )
    else:  # Claude
        api_key = st.text_input(
            "Anthropic Claude API Key",
            type="password",
            help="輸入您的 Anthropic API Key (https://console.anthropic.com)"
        )
    
    if api_key:
        st.session_state.api_key = api_key
        st.session_state.api_choice = api_choice
        st.success(f"✅ {api_choice} API Key 已載入")
    
    st.markdown("---")
    
        
    # 資料來源（自動載入）
    st.subheader("📊 資料來源")

    # 自動載入資料
    DATA_PATH = "all_types_wide_30_completed_sas.csv"

    if st.session_state.df_full is None:
        try:
            df_full = pd.read_csv(DATA_PATH)
            st.session_state.df_full = df_full
            
            # 提取可用屬性
            type_columns = [col.replace('_win_count', '').replace('_total_battles', '') 
                            for col in df_full.columns if '_win_count' in col]
            available_types = sorted(set(type_columns))
            st.session_state.available_types = available_types
        except FileNotFoundError:
            st.error(f"❌ 找不到資料檔案：{DATA_PATH}")

    if st.session_state.df_full is not None:
        st.success(f"✅ 資料已載入！共 {len(st.session_state.df_full)} 筆記錄")
        st.info(f"📋 可用屬性數量: {len(st.session_state.available_types)}")        
            
        
    
    # MCMC 參數設定
    st.subheader("🔬 蒙地卡羅馬可夫鏈 參數")
    
    n_warmup = st.number_input(
        "Burn-in 樣本數",
        min_value=500,
        max_value=5000,
        value=1000,
        step=100,
        help="Burn-in 樣本數"
    )
    
    n_samples = st.number_input(
        "Sampling 樣本數",
        min_value=1000,
        max_value=10000,
        value=2000,
        step=500,
        help="Sampling 樣本數"
    )
       
    n_chains = st.selectbox(
        "鏈數 (Chains)",
        options=[1, 2, 4],
        index=1,
        help="蒙地卡羅馬可夫鏈 數量"
    )
    
    target_accept = st.slider(
        "目標接受率 (Target Accept)",
        min_value=0.80,
        max_value=0.99,
        value=0.95,
        step=0.01,
        help="NUTS 採樣器的目標接受率，越高越謹慎但越慢"
    )    
    
    
    st.info(f"💡 總迭代數 = {n_warmup} (Warmup) + {n_samples} (Sample) = {n_warmup + n_samples}")
    st.markdown("---")
    
    # 關於系統
    with st.expander("ℹ️ 關於此系統"):
        st.markdown("""
        **貝氏後設分析系統**
        
        本系統使用貝氏階層模型進行後設分析，
        比較不同寶可夢屬性在多個道館中的對戰表現。
        
        **主要功能：**
        - 🎲 貝氏推論與後驗分佈
        - 📊 階層模型（借用資訊）
        - 📈 5 種視覺化圖表
        - 📋 完整統計報告
        - 🔍 收斂診斷
        
        **分析特色：**
        - 完整的 MCMC trace（warmup + posterior）
        - 中文 DAG 節點對照表
        - 預測新研究效果
        - 異質性評估
        """)

# 主要內容區
tab1, tab2, tab3 = st.tabs(["📊 貝氏分析", "📈 視覺化結果", "💬 AI 助手"])

# Tab 1: 貝氏分析
with tab1:
    st.header("📊 貝氏統合分析")
    
    if st.session_state.df_full is None:
        st.warning("⚠️ 資料載入失敗，請確認資料檔案存在")    
    else:
       
        # 顯示資料預覽
        with st.expander("👀 資料預覽"):
            st.dataframe(st.session_state.df_full, use_container_width=True)
        
        st.markdown("---")       
        
        
        # 屬性選擇
        st.subheader("🎯 選擇比較屬性")
        
        col1, col2 = st.columns(2)

        with col1:
            treatment_type = st.selectbox(
                "實驗組屬性（Treatment）",
                options=["請選擇..."] + st.session_state.available_types,
                index=0,
                help="選擇要分析的實驗組屬性"
            )

        with col2:
            control_type = st.selectbox(
                "對照組屬性（Control）",
                options=["請選擇..."] + st.session_state.available_types,
                index=0,
                help="選擇要分析的對照組屬性"
            )

        # 驗證選擇
        if treatment_type == "請選擇..." or control_type == "請選擇...":
            st.warning("⚠️ 請選擇實驗組和對照組屬性")
        elif treatment_type == control_type:
            st.error("❌ 實驗組和對照組不能相同")
        else:
            st.success(f"✅ 已選擇: {treatment_type} (實驗組) vs {control_type} (對照組)")        
        

            # 顯示資料摘要
            st.subheader("📋 資料摘要")
            
            df_full = st.session_state.df_full
            treatment_wins = df_full[f'{treatment_type}_win_count'].sum()
            treatment_total = df_full[f'{treatment_type}_total_battles'].sum()
            control_wins = df_full[f'{control_type}_win_count'].sum()
            control_total = df_full[f'{control_type}_total_battles'].sum()
            
            col1, col2, col3 = st.columns(3)
            
            with col1:
                st.metric(
                    "資料筆數",
                    f"{len(df_full)} 筆",
                    help="道館（研究單位）的數量"
                )
            
            with col2:
                treatment_rate = treatment_wins / treatment_total if treatment_total > 0 else 0
                st.metric(
                    f"{treatment_type} 平均勝率",
                    f"{treatment_rate:.3f}",
                    help="實驗組的整體勝率"
                )
            
            with col3:
                control_rate = control_wins / control_total if control_total > 0 else 0
                st.metric(
                    f"{control_type} 平均勝率",
                    f"{control_rate:.3f}",
                    help="對照組的整體勝率"
                )
            
            st.markdown("---")
            
            # 執行分析按鈕
            # 只有在正確選擇屬性時才顯示按鈕
            if treatment_type != "請選擇..." and control_type != "請選擇..." and treatment_type != control_type:
                if st.button("🔬 開始貝氏分析", type="primary", use_container_width=True):
                    # 執行分析...           
                    with st.spinner("🔄 正在執行 蒙地卡羅馬可夫鏈 抽樣，請稍候..."):
                        try:
                            # 初始化分析器
                            analyzer = BayesianMetaAnalyzer(
                                session_id=st.session_state.session_id,
                                treatment_type=treatment_type,
                                control_type=control_type
                            )
                            
                            # 載入資料
                            analyzer.load_data(
                                df_full=df_full,
                                treatment_type=treatment_type,
                                control_type=control_type
                            )
                            
                            # 執行分析
                            results = analyzer.run_analysis(
                                n_warmup=n_warmup,
                                n_samples=n_samples,
                                n_chains=n_chains,
                                target_accept=target_accept
                            )
                            
                            # 儲存結果
                            st.session_state.analysis_results = results
                            st.session_state.analyzer = analyzer
                            
                            st.success("✅ 分析完成！請切換到「視覺化結果」頁籤查看圖表")
                            st.balloons()
                            
                        except Exception as e:
                            st.error(f"❌ 分析失敗: {str(e)}")
                            st.exception(e)
                
                # 顯示結果（如果已執行分析）
                if st.session_state.analysis_results is not None:
                    results = st.session_state.analysis_results
                    
                    st.markdown("---")
                    st.subheader("📊 分析結果摘要")

                    overall = results['overall']

                    # 使用 metric 卡片顯示
                    col1, col2, col3 = st.columns(3)

                    with col1:
                        st.metric(
                            label="d (整體效應)",
                            value=f"{overall['d_mean']:.4f}",
                        )
                        st.markdown(f"<span style='color: green; font-size: 14px;'>↑ HDI: [{overall['d_hdi_low']:.3f}, {overall['d_hdi_high']:.3f}]</span>", unsafe_allow_html=True)

                    with col2:
                        st.metric(
                            label="勝算比 (OR)",
                            value=f"{overall['or_mean']:.3f}",
                        )
                        st.markdown(f"<span style='color: green; font-size: 14px;'>↑ HDI: [{overall['or_hdi_low']:.3f}, {overall['or_hdi_high']:.3f}]</span>", unsafe_allow_html=True)

                    with col3:
                        st.metric(
                            label="sigma (異質性)",
                            value=f"{overall['sigma_mean']:.4f}",
                        )
                        st.markdown(f"<span style='color: green; font-size: 14px;'>↑ HDI: [{overall['sigma_hdi_low']:.3f}, {overall['sigma_hdi_high']:.3f}]</span>", unsafe_allow_html=True)

                    # 異質性解釋
                    sigma_mean = overall['sigma_mean']
                    if sigma_mean > 0.5:
                        st.info("🔴 **高異質性**: 不同道館的結果差異很大")
                    elif sigma_mean > 0.3:
                        st.info("🟡 **中等異質性**: 不同道館的結果有一定差異")
                    else:
                        st.info("🟢 **低異質性**: 不同道館的結果相對一致")                    
                                      
                    st.markdown("---")
                    
                    # 預測效應
                    st.markdown("#### 🔮 預測新研究效果")
                    pred = results['predictive']
                    
                    col1, col2, col3 = st.columns(3)
                    
                    with col1:
                        st.metric(
                            "預測 delta_new",
                            f"{pred['delta_new_mean']:.4f}",
                            delta=f"± {pred['delta_new_sd']:.4f}",
                            help="預測第 31 間道館的效應值"
                        )
                    
                    with col2:
                        st.metric(
                            "預測 OR",
                            f"{pred['or_new_mean']:.4f}",
                            help="預測的勝算比"
                        )
                    
                    with col3:
                        st.metric(
                            "不確定性倍數",
                            f"{pred['uncertainty_ratio']:.2f}x",
                            help="相比整體效應的不確定性增加"
                        )
                    
                    st.markdown("---")
                    
                    # 收斂診斷
                    st.markdown("#### 🔍 模型診斷")
                    diag = results['diagnostics']
                    
                    col1, col2, col3 = st.columns(3)
                    
                    with col1:
                        rhat_status = "✅ 已收斂" if diag['converged'] else "❌ 未收斂"
                        st.metric("收斂狀態", rhat_status)
                    
                    with col2:
                        if diag['rhat_d'] is not None:
                            st.metric("R-hat (d)", f"{diag['rhat_d']:.4f}")
                    
                    with col3:
                        if diag['ess_d'] is not None:
                            st.metric("ESS (d)", f"{int(diag['ess_d'])}")
                    
                    # 下載報告
                    st.markdown("---")
                    st.markdown("#### 📥 下載報告")
                    
                    report_text = format_summary_stats(results)
                    
                    st.download_button(
                        label="📄 下載完整分析報告 (.txt)",
                        data=report_text,
                        file_name=f"meta_analysis_{treatment_type}_vs_{control_type}_{datetime.now().strftime('%Y%m%d_%H%M%S')}.txt",
                        mime="text/plain",
                        use_container_width=True
                    )

# Tab 2: 視覺化結果
with tab2:
    st.header("📈 視覺化結果")
    
    if st.session_state.analysis_results is None:
        st.info("ℹ️ 請先在「貝氏分析」頁面執行分析")
    else:
        results = st.session_state.analysis_results
        analyzer = st.session_state.analyzer
        
        # 創建 5 個 tab
        viz_tabs = st.tabs([
            "🔍 模型結構 (DAG)",
            "📋 DAG 對照表", 
            "📉 Trace Plot",
            #"📊 Posterior Plot",
            "🌲 Forest Plot"
        ])
        
        # Tab 1: DAG
        with viz_tabs[0]:
            st.markdown("### 🔍 模型結構圖 (DAG)")
            st.markdown("""
            **DAG（Directed Acyclic Graph）** 顯示貝氏階層模型的結構：
            - **橢圓節點**：隨機變數（prior 和 likelihood）
            - **矩形節點**：觀測資料
            - **菱形節點**：確定性變數（由其他變數推導）
            - **箭頭**：依賴關係
            """)
            
            with st.spinner("正在生成 DAG..."):
                try:
                    dag_img = plot_dag(analyzer.model, results)
                    if dag_img:
                        st.image(dag_img, use_column_width=True)
                    else:
                        st.warning("⚠️ 無法生成 DAG 圖（可能需要安裝 Graphviz）")
                except Exception as e:
                    st.error(f"生成 DAG 失敗: {str(e)}")
        
        # Tab 2: DAG 對照表
        with viz_tabs[1]:
            st.markdown("### 📋 DAG 中文對照表")
            st.markdown("節點說明表格：解釋 DAG 中每個節點的統計意義和實際情境應用")
            
            from meta_analysis_utils import create_dag_legend_html
            dag_html = create_dag_legend_html(results)
            st.markdown(dag_html, unsafe_allow_html=True)        
        
       
        # Tab 3: Trace Plot
        with viz_tabs[2]:
            st.markdown("### 📉 Trace Plot（抽樣軌跡）")
            st.markdown("""
            **Trace Plot 用途**：
            - **左圖**：後驗分佈的密度圖
            - **右圖**：完整的 MCMC 軌跡（warmup + posterior）
            - **紅色虛線**：標記 burn-in 結束點
            
            **判斷收斂**：
            - 軌跡圖應該要穩定震盪
            - 不同鏈應該混合良好
            - R-hat < 1.1 表示已收斂
            """)
            
            with st.spinner("正在生成 Trace Plot..."):
                try:
                    trace_img = plot_trace_combined(results)
                    if trace_img:
                        st.image(trace_img, use_column_width=True)
                except Exception as e:
                    st.error(f"生成 Trace Plot 失敗: {str(e)}")
        
        # Tab 4: Posterior Plot

        # Tab 5: Forest Plot
        #with viz_tabs[4]:
        with viz_tabs[3]:    
            st.markdown("### 🌲 Forest Plot（各道館效應）")
            st.markdown("""
            **Forest Plot 用途**：
            - 顯示每個道館的特定效應（delta）
            - **點**：平均效應值
            - **橫線**：95% HDI
            - **垂直虛線**：零效應參考線
            
            **解讀**：
            - HDI 不包含 0 → 該道館有顯著效應
            - 點在右側（> 0）→ 實驗組優勢
            - 點在左側（< 0）→ 對照組優勢
            """)
            
            with st.spinner("正在生成 Forest Plot..."):
                try:
                    forest_img = plot_forest(results)
                    if forest_img:
                        st.image(forest_img, use_column_width=True)
                except Exception as e:
                    st.error(f"生成 Forest Plot 失敗: {str(e)}")

# Tab 3: AI 助手
with tab3:
    st.header("💬 AI 分析助手")
    
    if not st.session_state.get('api_key'):
        st.warning("⚠️ 請在左側輸入您的 API Key 以使用 AI 助手")
        st.info("💡 支援 Google Gemini 和 Anthropic Claude")
    elif st.session_state.analysis_results is None:
        st.info("ℹ️ 請先在「貝氏分析」頁面執行分析")
    else:
        # 初始化 LLM 助手
        if st.session_state.llm_assistant is None:
            api_choice = st.session_state.get('api_choice', 'Google Gemini')
            st.session_state.llm_assistant = MetaAnalysisLLMAssistant(
                api_key=st.session_state.api_key,
                session_id=st.session_state.session_id,
                api_provider=api_choice
            )
        
        # 聊天容器
        chat_container = st.container()
                        
        with chat_container:
            for message in st.session_state.chat_history:
                with st.chat_message(message["role"]):
                    st.markdown(message["content"])
        
        # 使用者輸入
        if prompt := st.chat_input("詢問關於分析結果的任何問題..."):
            # 添加使用者訊息
            st.session_state.chat_history.append({
                "role": "user",
                "content": prompt
            })
            
            with st.chat_message("user"):
                st.markdown(prompt)
            
            # AI 回應
            with st.chat_message("assistant"):
                with st.spinner("思考中..."):
                    try:
                        response = st.session_state.llm_assistant.get_response(
                            user_message=prompt,
                            analysis_results=st.session_state.analysis_results
                        )
                        st.markdown(response)
                        
                    except Exception as e:
                        error_msg = f"❌ 錯誤: {str(e)}\n\n請檢查 API key 或重新表達問題。"
                        st.error(error_msg)
                        response = error_msg
            
            # 添加助手回應
            st.session_state.chat_history.append({
                "role": "assistant",
                "content": response
            })
        
        st.markdown("---")
        
        # 快速問題按鈕
        st.subheader("💡 快速問題")
        
        quick_questions = [
            "📊 給我這次分析的總結",
            "🎯 解釋 d 和勝算比",
            "🔍 解釋 sigma（異質性）",
            "❓ 什麼是貝氏統合分析？",
            "🔮 解釋預測新道館",
            "🆚 貝氏 vs 頻率論",
            "⚔️ 對戰策略建議",
            "🏛️ 比較不同道館"
        ]
        
        cols = st.columns(4)
        for idx, question in enumerate(quick_questions):
            col_idx = idx % 4
            if cols[col_idx].button(question, key=f"quick_{idx}"):
                # 根據問題選擇對應的方法
                if "總結" in question:
                    response = st.session_state.llm_assistant.generate_summary(
                        st.session_state.analysis_results
                    )
                elif "d 和勝算比" in question:
                    response = st.session_state.llm_assistant.explain_metric(
                        'd',
                        st.session_state.analysis_results
                    )
                elif "sigma" in question or "異質性" in question:
                    response = st.session_state.llm_assistant.explain_metric(
                        'sigma',
                        st.session_state.analysis_results
                    )
                elif "貝氏後設分析" in question:
                    response = st.session_state.llm_assistant.explain_bayesian_meta_analysis()
                elif "預測" in question:
                    response = st.session_state.llm_assistant.explain_predictive_inference(
                        st.session_state.analysis_results
                    )
                elif "貝氏" in question and "頻率論" in question:
                    response = st.session_state.llm_assistant.explain_bayesian_meta_analysis()
                elif "策略" in question:
                    response = st.session_state.llm_assistant.battle_strategy_advice(
                        st.session_state.analysis_results
                    )
                elif "道館" in question:
                    response = st.session_state.llm_assistant.compare_gyms(
                        st.session_state.analysis_results
                    )
                else:
                    response = st.session_state.llm_assistant.get_response(
                        question,
                        st.session_state.analysis_results
                    )
                
                # 添加到聊天歷史
                st.session_state.chat_history.append({
                    "role": "user",
                    "content": question
                })
                
                st.session_state.chat_history.append({
                    "role": "assistant",
                    "content": response
                })
                
                st.rerun()
        
        # 重置對話按鈕
        st.markdown("---")
        if st.button("🔄 重置對話"):
            st.session_state.llm_assistant.reset_conversation()
            st.session_state.chat_history = []
            st.success("✅ 對話已重置")
            st.rerun()

# Footer
st.markdown("---")
st.markdown(
    f"""
    <div style='text-align: center'>
        <p>🎲 Bayesian Meta-Analysis System for Pokémon Type Comparison | Built with Streamlit & PyMC</p>
        <p>Session ID: {st.session_state.session_id[:8]}</p>
    </div>
    """,
    unsafe_allow_html=True
)