472_问答系统设计¶

# ---- 方案1：抽取式QA（BERT在原文中定位答案）----
from transformers import pipeline

# 加载QA流水线
qa_pipeline = pipeline(
    "question-answering",
    model="deepset/roberta-base-squad2"
)

context = """
BERT (Bidirectional Encoder Representations from Transformers) was introduced
by Google in 2018. It uses a bidirectional Transformer encoder and is pre-trained
on masked language modeling and next sentence prediction tasks.
"""
question = "When was BERT introduced?"

result = qa_pipeline(question=question, context=context)
print(f"答案: {result['answer']}")        # 2018
print(f"置信度: {result['score']:.4f}")   # 对应位置的概率

# ---- 方案2：RAG系统（检索+生成）----
# pip install langchain chromadb sentence-transformers

from langchain.vectorstores import Chroma
from langchain.embeddings import HuggingFaceEmbeddings
from langchain.schema import Document

# 2a. 构建向量知识库
embedding_model = HuggingFaceEmbeddings(
    model_name='sentence-transformers/all-MiniLM-L6-v2'  # 轻量embedding模型
)

# 模拟文档库（实际从PDF/数据库加载）
docs = [
    Document(page_content="BERT于2018年由Google发布，基于双向Transformer。"),
    Document(page_content="GPT系列由OpenAI开发，使用自回归语言模型。"),
    Document(page_content="T5把所有NLP任务统一成文本到文本的格式。"),
]

# 构建向量数据库（FAISS或Chroma）
vectorstore = Chroma.from_documents(
    docs,
    embedding=embedding_model,
    persist_directory="./chroma_db"  # 持久化到磁盘
)

# 2b. 检索最相关文档
retriever = vectorstore.as_retriever(
    search_kwargs={"k": 2}  # 返回最相似的2个文档
)
query = "BERT是什么时候发布的？"
retrieved_docs = retriever.get_relevant_documents(query)
for doc in retrieved_docs:
    print("检索到:", doc.page_content)

# 2c. 用LLM生成最终答案（需配置API）
from langchain.llms import OpenAI
from langchain.chains import RetrievalQA

llm = OpenAI(temperature=0)  # temperature=0 减少随机性
qa_chain = RetrievalQA.from_chain_type(
    llm=llm,
    chain_type="stuff",     # 把所有检索文档塞进prompt
    retriever=retriever
)
answer = qa_chain.run(query)
print("RAG答案:", answer)

# ---- 方案3：BERT-QA的span预测原理演示 ----
from transformers import BertForQuestionAnswering, BertTokenizer
import torch

tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
model = BertForQuestionAnswering.from_pretrained('bert-large-uncased-whole-word-masking-finetuned-squad')

q = "What is BERT?"
c = "BERT is a language model developed by Google in 2018."

# 问题和上下文一起编码，用[SEP]分隔
enc = tokenizer(q, c, return_tensors='pt')
with torch.no_grad():
    outputs = model(**enc)
    # start_logits 和 end_logits 对每个token位置打分
    start_idx = outputs.start_logits.argmax()   # 答案起始token位置
    end_idx = outputs.end_logits.argmax()       # 答案终止token位置

tokens = tokenizer.convert_ids_to_tokens(enc['input_ids'][0])
answer_tokens = tokens[start_idx: end_idx + 1]
print("答案tokens:", tokenizer.convert_tokens_to_string(answer_tokens))