app.py

import gradio as gr
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from PyPDF2 import PdfReader
import google.generativeai as genai
import os
from langsmith import Client
from ragas.metrics import faithfulness, answer_relevancy, context_relevancy

# 更新的 langchain_community 導入
from langchain_community.llms import OpenAI  # 示例導入

# 加載模型
openelm_model = AutoModelForCausalLM.from_pretrained(
    "apple/OpenELM-270M", 
    trust_remote_code=True
)

# 加載 tokenizer，確保 trust_remote_code=True
openelm_tokenizer = AutoTokenizer.from_pretrained(
    "apple/OpenELM-270M", 
    trust_remote_code=True
)

# 設置 Gemini API
GOOGLE_API_KEY = os.getenv("GOOGLE_API_KEY")
genai.configure(api_key=GOOGLE_API_KEY)

# 設置 LangSmith
os.environ["LANGCHAIN_API_KEY"] = "your_langchain_api_key"
os.environ["LANGCHAIN_TRACING_V2"] = "true"
os.environ["LANGCHAIN_ENDPOINT"] = "https://api.smith.langchain.com"
client = Client()

def extract_text_from_pdf(pdf_path):
    with open(pdf_path, 'rb') as file:
        reader = PdfReader(file)
        text = ""
        for page in reader.pages:
            text += page.extract_text() + "\n"
    return text

def gemini_generate(prompt, max_tokens):
    model = genai.GenerativeModel('gemini-pro')
    response = model.generate_content(prompt, max_output_tokens=max_tokens)
    return response.text

def openelm_generate(prompt, max_tokens):
    tokenized_prompt = openelm_tokenizer(prompt, return_tensors="pt")
    output_ids = openelm_model.generate(
        tokenized_prompt["input_ids"],
        max_length=max_tokens,
        pad_token_id=0,
    )
    return openelm_tokenizer.decode(output_ids[0], skip_special_tokens=True)

def evaluate_response(response, context, query):
    faith_score = faithfulness.score([response], [context], [query])
    ans_rel_score = answer_relevancy.score([response], [query])
    ctx_rel_score = context_relevancy.score([response], [context], [query])
    return faith_score, ans_rel_score, ctx_rel_score

def process_query(pdf_file, llm_choice, query, max_tokens, api_key):
    try:
        global GOOGLE_API_KEY
        if api_key:
            GOOGLE_API_KEY = api_key
            genai.configure(api_key=GOOGLE_API_KEY)
        
        # 從 PDF 提取文本
        pdf_path = pdf_file.name
        context = extract_text_from_pdf(pdf_path)
        
        # 根據選擇的 LLM 生成回應
        if llm_choice == "Gemini":
            response = gemini_generate(f"上下文: {context}\n問題: {query}", max_tokens)
        else:  # OpenELM
            response = openelm_generate(f"上下文: {context}\n問題: {query}", max_tokens)
        
        # 評估回應
        faith_score, ans_rel_score, ctx_rel_score = evaluate_response(response, context, query)
        
        return response, faith_score, ans_rel_score, ctx_rel_score
    except Exception as e:
        return str(e), 0, 0, 0  # 返回錯誤消息和零分數

# Gradio 介面
iface = gr.Interface(
    fn=process_query,
    inputs=[
        gr.File(label="上傳 PDF"),
        gr.Dropdown(["Gemini", "OpenELM"], label="選擇 LLM"),
        gr.Textbox(label="輸入您的問題"),
        gr.Slider(minimum=50, maximum=1000, step=50, label="最大令牌數"),
        gr.Textbox(label="Gemini API 金鑰 (可選)", type="password")
    ],
    outputs=[
        gr.Textbox(label="生成的答案"),
        gr.Number(label="真實性得分"),
        gr.Number(label="答案相關性得分"),
        gr.Number(label="上下文相關性得分")
    ],
    title="多模型 LLM 查詢介面，支持 PDF 上下文",
    description="上傳 PDF，選擇 LLM，並提出問題。回應將使用 RAGAS 指標進行評估。"
)

if __name__ == "__main__":
    iface.launch()