app.py

import gradio as gr
import fitz  # PyMuPDF
from transformers import pipeline
from keybert import KeyBERT
from sentence_transformers import SentenceTransformer
import faiss
import numpy as np
import re

# =============== 初始化模型 ===============
summarizer = pipeline("summarization", model="facebook/bart-large-cnn")
qa_generator = pipeline("text2text-generation", model="google/flan-t5-base")
embedder = SentenceTransformer("sentence-transformers/all-MiniLM-L6-v2")
kw_model = KeyBERT(model=embedder)

# 全局变量（保存上传文献的分块与索引）
CHUNKS = []
INDEX = None


# =============== 工具函数 ===============
def clean_text(t):
    t = t.replace("\x00", " ")
    t = re.sub(r"\s+", " ", t)
    return t.strip()


def pdf_to_text(file_bytes):
    doc = fitz.open(stream=file_bytes, filetype="pdf")
    texts = []
    for page in doc:
        t = page.get_text("text")
        if t:
            texts.append(t)
    return clean_text("\n".join(texts))


def chunk_text(text, chunk_size=800, overlap=120):
    if len(text) <= chunk_size:
        return [text]
    chunks, start = [], 0
    while start < len(text):
        end = min(start + chunk_size, len(text))
        boundary = max(
            text.rfind(". ", start, end),
            text.rfind("。", start, end),
            text.rfind("\n", start, end),
        )
        if boundary == -1 or boundary <= start + 200:
            boundary = end
        chunks.append(text[start:boundary].strip())
        start = max(boundary - overlap, 0)
        if start == boundary:
            start += 1
    return [c for c in chunks if len(c) > 10]


def build_faiss(chunks):
    global INDEX
    embs = embedder.encode(chunks, normalize_embeddings=True)
    index = faiss.IndexFlatIP(embs.shape[1])
    index.add(embs.astype(np.float32))
    INDEX = index


def retrieve(query, top_k=5):
    if INDEX is None or not CHUNKS:
        return []
    q = embedder.encode([query], normalize_embeddings=True).astype(np.float32)
    D, I = INDEX.search(q, top_k)
    results = [(CHUNKS[i], float(D[0][j])) for j, i in enumerate(I[0]) if i != -1]
    return results


# =============== 功能实现 ===============
def handle_upload(pdf_files):
    global CHUNKS
    if not pdf_files:
        return "", "未上传文件"
    texts = []
    for f in pdf_files:
        b = f.read()
        txt = pdf_to_text(b)
        texts.append(txt)
    merged = "\n".join(texts)
    CHUNKS = chunk_text(merged)
    build_faiss(CHUNKS)
    return merged[:50000], f"✅ 已成功建立索引：{len(CHUNKS)} 个片段"


def summarize_text(text):
    if not text or len(text) < 50:
        return "请输入更长的文本或先上传 PDF。"
    parts = chunk_text(text, 1000, 100)
    summaries = []
    for p in parts:
        try:
            s = summarizer(p, max_length=150, min_length=40, do_sample=False)[0]["summary_text"]
            summaries.append(s)
        except Exception:
            continue
    combined = " ".join(summaries)
    final = summarizer(combined, max_length=200, min_length=60, do_sample=False)[0]["summary_text"]
    return final


def extract_keywords(text, top_n=10):
    if not text or len(text) < 50:
        return "请输入更长的文本或先上传 PDF。"
    pairs = kw_model.extract_keywords(text[:10000], top_n=top_n)
    return ", ".join([k for k, _ in pairs])


def answer_question(question, top_k=5):
    if not CHUNKS:
        return "请先上传 PDF 并建立索引。", ""
    docs = retrieve(question, top_k)
    context = "\n\n".join([f"[{i+1}] {c}" for i, (c, _) in enumerate(docs)])
    prompt = (
        "You are a helpful research assistant. Answer the question strictly based on the CONTEXT. "
        "If the answer cannot be found, say 'Not found in the provided documents.'\n\n"
        f"CONTEXT:\n{context}\n\nQUESTION: {question}\nANSWER:"
    )
    out = qa_generator(prompt, max_new_tokens=256)[0]["generated_text"]
    cites = "\n".join([f"[{i+1}] 相似度={score:.3f}" for i, (_, score) in enumerate(docs)])
    return out, cites


# =============== Gradio 界面 ===============
with gr.Blocks(title="Paper Reader Assistant") as demo:
    gr.Markdown("""
    # 📖 Paper Reader Assistant
    上传 PDF，自动抽取文本，生成摘要、关键词，并支持基于内容的问答（RAG）。
    """)

    with gr.Row():
        pdf_uploader = gr.File(label="上传 PDF（可多选）", file_count="multiple", file_types=[".pdf"])
        build_info = gr.Textbox(label="状态", interactive=False)

    # 修改点 1：使用 Textbox 替代 Dataframe
    with gr.Row():
        doc_text = gr.Textbox(label="文档全文（前 50,000 字符预览）", lines=14)
        file_table = gr.Textbox(label="文件状态", lines=3)

    upload_btn = gr.Button("📥 解析 PDF 并建立索引")

    with gr.Tab("📝 摘要"):
        sum_btn = gr.Button("生成摘要")
        sum_out = gr.Textbox(label="摘要结果", lines=10)

    with gr.Tab("🔑 关键词"):
        kw_btn = gr.Button("提取关键词")
        kw_out = gr.Textbox(label="关键词", lines=8)

    with gr.Tab("❓ 问答 RAG"):
        question = gr.Textbox(label="你的问题", lines=2)
        qa_btn = gr.Button("回答问题")
        answer_out = gr.Textbox(label="答案", lines=10)
        cites_out = gr.Textbox(label="参考片段", lines=6)

    # 修改点 2：绑定时删除 Dataframe 输出
    upload_btn.click(handle_upload, inputs=[pdf_uploader], outputs=[doc_text, build_info])
    sum_btn.click(summarize_text, inputs=[doc_text], outputs=sum_out)
    kw_btn.click(extract_keywords, inputs=[doc_text], outputs=kw_out)
    qa_btn.click(answer_question, inputs=[question], outputs=[answer_out, cites_out])

# 修改点 3：允许远程访问、生成共享链接
if __name__ == "__main__":
    demo.launch(server_name="0.0.0.0", server_port=7860, share=True)