app.py

import os
import sys
import math
from openai import OpenAI
import requests
import gradio as gr
import pandas as pd
import concurrent.futures
from datasets import Dataset
from tqdm import tqdm
from ragas import evaluate, SingleTurnSample
from ragas.llms import LangchainLLMWrapper
from ragas.embeddings import LangchainEmbeddingsWrapper
from langchain_openai import ChatOpenAI, OpenAIEmbeddings
from ragas.metrics import (
    ResponseRelevancy, LLMContextPrecisionWithReference, LLMContextRecall,
    ContextEntityRecall, Faithfulness, NoiseSensitivity, SemanticSimilarity, FactualCorrectness
)

# 設定輸出編碼為 UTF-8（解決中文顯示問題）
sys.stdout.reconfigure(encoding="utf-8")

# 從Google Drive下載 Ground Truth
gt_url = os.environ.get("GT_URL")
gt_path = "tender_groundtruth.csv"

if gt_url and not os.path.exists(gt_path):
    print("嘗試下載 Ground Truth...")
    r = requests.get(gt_url)
    print("HTTP 狀態碼：", r.status_code)
    if r.status_code != 200:
        print("下載失敗內容預覽：", r.text[:500])
    else:
        with open(gt_path, "wb") as f:
            f.write(r.content)

# 綁定實驗室Google帳號(Python TA)Google Sheet，以記錄評估logs
def log_to_google_sheet(question, answer, contexts, scores):
    url = os.environ.get("G_SHEET_URL")
    if not url:
        print("G_SHEET_URL 未設定，略過記錄")
        return
    try:
        payload = {
            "question": question,
            "answer": answer,
            "contexts": contexts,
            "faithfulness": scores.get("Faithfulness"),
            "answer_relevancy": scores.get("Answer Relevancy"),
            "semantic_similarity": scores.get("Semantic Similarity"),
            "context_precision": scores.get("Context Precision"),
            "context_recall": scores.get("Context Recall"),
            "context_entity_recall": scores.get("Context Entity Recall")
        }
        response = requests.post(url, json=payload)
        print("成功寫入 Google Sheet:", response.status_code)
    except Exception as e:
        print("寫入 Google Sheet 失敗:", str(e))

def fetch_announcement_from_sheet():
    DEFAULT_MESSAGE = "尚無公告"
    try:
        url = os.environ.get("ANNOUNCEMENT_URL")
        if not url:
            print("Warning: 環境變數 'ANNOUNCEMENT_URL' 未設定")
            return DEFAULT_MESSAGE

        df = pd.read_csv(url)

        if "Announcement" not in df.columns:
            print("Error: CSV 檔案中無 'Announcement' 欄位")
            return DEFAULT_MESSAGE

        content = str(df["Announcement"].iloc[0]).strip()
        content = content.replace("\\n", "<br>").replace("\n", "<br>")
        return content if content else DEFAULT_MESSAGE

    except Exception as e:
        print(f"Error: 載入公告失敗：{e}")
        return DEFAULT_MESSAGE


def validate_openai_key(api_key):
    try:
        client = OpenAI(api_key=api_key)
        client.chat.completions.create(
            model="gpt-3.5-turbo",
            messages=[{"role": "user", "content": "hi"}],
            max_tokens=1
        )
        return None
    except Exception as e:
        err_msg = str(e)
        if "Incorrect API key provided" in err_msg:
            return pd.DataFrame([{"錯誤訊息": " 您輸入的 OpenAI API Key 有誤，請確認是否貼錯、字數不符或格式異常。"}]), None
        elif "exceeded your current quota" in err_msg:
            return pd.DataFrame([{"錯誤訊息": "您的 OpenAI 帳戶額度已用盡，請前往帳戶頁面檢查餘額。"}]), None
        elif "Rate limit" in err_msg:
            return pd.DataFrame([{"錯誤訊息": "OpenAI 請求頻率過高，請稍後再試"}]), None
        else:
            return pd.DataFrame([{"錯誤訊息": f"API Key 錯誤：{err_msg}"}]), None
    
def RAG_evaluation(uploaded_file, user_api_key):
    try:
        # 檢查 OpenAI API Key 是否有效
        validation_result = validate_openai_key(user_api_key)
        if validation_result:
            return validation_result
        
        os.environ["OPENAI_API_KEY"] = user_api_key
        print("評估開始")

        if not os.path.exists(gt_path):
            print("找不到 Ground Truth！")
            return pd.DataFrame(), None

        gt_df = pd.read_csv(gt_path)
        df = pd.read_csv(uploaded_file.name, converters={"Context": eval})
        print(f"上傳檔案筆數：{len(df)}，GT 檔案筆數：{len(gt_df)}")

        merged_df = pd.merge(df, gt_df[["Question", "Answer"]], on="Question", suffixes=("", "_GroundTruth"))
        merged_df = merged_df.rename(columns={"Answer_GroundTruth": "GroundTruth"})
        print(f"成功合併筆數：{len(merged_df)} / {len(df)}")
        if len(merged_df) < len(df):
            missing = df[~df["Question"].isin(merged_df["Question"])]
            print("未合併題目：", missing["Question"].tolist())
        if merged_df.empty:
            return pd.DataFrame([{"錯誤訊息": "合併後無資料，請確認題目與 GT 是否對應"}]), None
        
        llm_wrapper = LangchainLLMWrapper(ChatOpenAI(model="gpt-4o-mini-2024-07-18"))
        embedding_wrapper = LangchainEmbeddingsWrapper(OpenAIEmbeddings(model="text-embedding-3-large"))
        
        batch_size = 10
        records = []
        for batch_start in tqdm(range(0, len(merged_df), batch_size), desc="RAGAS Batch Evaluating"):
            batch_df = merged_df.iloc[batch_start:batch_start + batch_size]

            samples = []
            for _, row in batch_df.iterrows():
                if not isinstance(row["Context"], list):
                    print(f"Context 非 list，跳過。值：{row['Question']}")
                    continue

                sample = SingleTurnSample(
                    user_input=row["Question"],
                    response=row["Answer"],
                    retrieved_contexts=row["Context"],
                    reference=row["GroundTruth"],
                )
                samples.append(sample)
        
            try:
                dataset = Dataset.from_list([s.to_dict() for s in samples])
                result = evaluate(
                    dataset=dataset,
                    metrics=[
                        LLMContextPrecisionWithReference(),     # context precision
                        LLMContextRecall(),                     # context recall     
                        ContextEntityRecall(),
                        # NoiseSensitivity(), 
                        Faithfulness(),                         # faithfulness
                        ResponseRelevancy(),                    # answer relevancy
                        SemanticSimilarity(),                   # semantic similarity
                        # FactualCorrectness()
                    ],
                    llm=llm_wrapper,
                    embeddings=embedding_wrapper,
                    show_progress=True
                )

                result_df = result.to_pandas()

                for i, row in enumerate(result_df.itertuples()):
                    input_row = batch_df.iloc[i]
                    record = {
                        "Question": input_row["Question"],
                        "Faithfulness": getattr(row, "faithfulness", None),
                        "Answer Relevancy": getattr(row, "answer_relevancy", None),
                        "Semantic Similarity": getattr(row, "semantic_similarity", None),
                        # "Factual Correctness": getattr(row, "factual_correctness", None),
                        "Context Precision": getattr(row, "llm_context_precision_with_reference", None),
                        "Context Recall": getattr(row, "context_recall", None),
                        "Context Entity Recall": getattr(row, "context_entity_recall", None),
                        # "Noise Sensitivity": getattr(row, "noise_sensitivity_relevant", None)
                    }

                    for key in list(record.keys()):
                        val = record[key]
                        if isinstance(val, float) and not math.isfinite(val):
                            record[key] = ""
                
                    records.append(record)

                    log_to_google_sheet(
                        question=input_row["Question"],
                        answer=input_row["Answer"],
                        contexts=input_row["Context"],
                        scores=record
                    )

            except Exception as e:
                print(f"批次評估失敗（第 {batch_start+1} 筆起）：{e}")
                continue

        score_df = pd.DataFrame(records).fillna("")
        print("完成評估筆數：", len(score_df))
        
        numeric_cols = score_df.drop(columns=["Question"]).select_dtypes(include="number")
        if not numeric_cols.empty:
            avg_row = numeric_cols.mean().to_dict()
            avg_row["Question"] = "Average"
            score_df = pd.concat([score_df, pd.DataFrame([avg_row])], ignore_index=True)

        original_name = os.path.basename(uploaded_file.name)
        filename = os.path.splitext(original_name)[0]
        output_path = f"{filename}_result.csv"
        score_df.to_csv(output_path, index=False, encoding="utf-8-sig")
        print("評估結果已儲存：", output_path)

        return score_df, output_path

    except Exception as e:
        print("評估函式整體錯誤：", str(e))
        return pd.DataFrame([{"錯誤訊息": f"系統錯誤：{str(e)}"}]), None

# handle exception並執行RAG評估
def check_csv_and_run(file, key):
    if file is None:
        return pd.DataFrame([{"錯誤訊息": "請上傳檔案！"}]), None
    
    if not key or key.strip() == "":
        return pd.DataFrame([{"錯誤訊息": "請輸入 OpenAI API Key"}]), None

    try:
        df = pd.read_csv(file.name, encoding="utf-8-sig")
        df.columns = [col.strip() for col in df.columns]

        required_columns = {"Question", "Context", "Answer"}
        actual_columns = set(df.columns)

        if actual_columns != required_columns:
            return pd.DataFrame([{"錯誤訊息": f"欄位錯誤：應包含欄位 {required_columns}，實際為 {actual_columns}"}]), None

        if df.shape[0] == 0:
            return pd.DataFrame([{"錯誤訊息": "檔案中沒有資料列!"}]), None

        invalid_rows = df[df["Question"].notnull() & (df["Answer"].isnull() | df["Context"].isnull())]
        if len(invalid_rows) > 0:
            missing_questions = "\n".join(f"- {q}" for q in invalid_rows["Question"].tolist())
            return pd.DataFrame([{"錯誤訊息": f"發現 {len(invalid_rows)} 筆資料中 Answer 或 Context 為空：\n{missing_questions}"}]), None
        
        # check eval context
        try:
            for i, val in df["Context"].dropna().items():
                if not isinstance(eval(val), list):
                    return pd.DataFrame([{"錯誤訊息": f"第 {i + 1} 筆 Context 欄格式錯誤，請確認其內容應為 list"}]), None
        except Exception as e:
            return pd.DataFrame([{"錯誤訊息": f"Context 欄格式解析錯誤，請確認其為有效的 list 格式，例如 ['A', 'B']：{str(e)}"}]), None
        
        # 若上傳之待評估檔案無錯誤，執行評估
        try:
            return RAG_evaluation(file, key)
        # 檢查 OpenAI API Key 是否有效
        except Exception as e:
            error_message = str(e)
            return pd.DataFrame([{"錯誤訊息": f"系統錯誤：{error_message}"}]), None
    except Exception as e:
        return pd.DataFrame([{"錯誤訊息": f"評估失敗：{str(e)}"}]), None

# Gradio 介面
with gr.Blocks() as demo:
    gr.Markdown("""
# 📐 RAG系統評估工具 (分流A)
                
### 📄 使用說明
- 請上傳您 RAG 系統產出的結果檔案（需包含欄位：Question、Context、Answer），並填入您的 OpenAI API Key，以進行評估。
- ⏳ 完整評估**通常需耗時 1 小時以上**，若無即時回應，請**耐心等候**，系統並未當機，謝謝您的理解。

### 🚦 分流措施
本工具部署於 Hugging Face Public Space，若同時有多位使用者使用，系統會將您的評估請求**排入佇列**。
為避免長時間等待，建議您**先僅送出 1 筆資料進行測試**，若進度條顯示之預估**等待時間超過 2 小時（7000 秒以上），可能是其他使用者正在使用**。
本頁為**分流 A**，您可以考慮改用其他分流或稍後再試，感謝您的耐心與配合!

- 🔁 [主頁面 (Main)](https://huggingface.co/spaces/KSLab/RAG_Evaluator)
- 🔁 [分流 A](https://huggingface.co/spaces/KSLab/RAG_Evaluator_A)
- 🔁 [分流 C](https://huggingface.co/spaces/KSLab/RAG_Evaluator_C)

### 📢 系統公告
""")
    gr.Markdown(fetch_announcement_from_sheet())

    file_input = gr.File(label="上傳 Evaluation_Dataset.csv")
    api_key_input = gr.Textbox(label="OpenAI API Key", type="password")
    submit_btn = gr.Button("開始評估")

    result_output = gr.Dataframe(label="評估結果")
    download_link = gr.File(label="下載評估結果(CSV)")

    # 常見QA文字
    gr.Markdown("""
    ---  
    ### ❓ 常見問題 & 解答  

    **Q: 什麼是「指令集」？**  
    A: 「指令集」是我們用來描述老師在課堂上所設計的各種學習活動操作流程。在與教學系統互動時，老師通常會透過一系列結構化的指令來引導學生完成任務，因此我們將這些可重複使用的操作流程統稱為「指令集」。
                指令集也如同RESTful API一樣，我們有先盡力的與老師們溝通他們的需求，不過這些需求都只能視為一個草案，最終仍需要仰賴得標業者與老師們收斂，並且確定最終的版本來加以實作。
    """)

    def wrapped_fn(file, key):
        return RAG_evaluation(file, key)

    submit_btn.click(
        fn=check_csv_and_run,
        inputs=[file_input, api_key_input],
        outputs=[result_output, download_link],
)

demo.launch()