1M 上下文窗口 LLM 2026 实测：大模型的潜力与局限

# 随机抽取 100 条早期信息
accuracy = sum(
    model.recall(question, answer) 
    for question, answer in early_questions
) / 100

book = load_book("war_and_peace.txt")  # 1.2M 字符
summary = llm.summarize(book)
# 输出：500 字摘要

# 索引整个 GitHub 仓库
rg "def " --type py | head -10000 > functions.txt

# 询问
"这个项目中哪些函数处理用户认证？"

df = pd.read_csv("large_dataset.csv")  # 1M 行

question = "2023 年增长最快的 5 个产品线？"
answer = llm.query(df, question)

from langchain.vectorstores import Chroma
from langchain.embeddings import OpenAIEmbeddings

# 将大文档切块向量化
chunks = chunker.split(document)
db = Chroma.from_documents(chunks, embeddings)

# 先检索相关块
relevant = db.similarity_search(question, k=10)

from langchain.chains import LLMChain
from langchain.prompts import PromptTemplate

compress_prompt = PromptTemplate.from_template(
    "压缩以下文本，保留关键信息：{text}"
)
compressed = chain.run(text=long_text)

# 注意力分布不均
attention_weights = model.get_attention_weights()

# 前 10% 内容注意力 ≪ 后 90%

from unstructured import partition

# 文档解析
elements = partition(filename="document.pdf")

# 分块 + 向量化
chunks = chunk_elements(elements)
embeddings = vectorize(chunks)

from langchain_community.document_loaders import PyPDFLoader
from langchain.text_splitter import RecursiveCharacterTextSplitter

loader = PyPDFLoader("large.pdf")
docs = loader.load()

splitter = RecursiveCharacterTextSplitter(
    chunk_size=10000,
    chunk_overlap=1000
)

contract = load_contract("master_agreement.pdf")  # 500 页

clauses = llm.extract_clauses(contract, [
    "付款条款",
    "违约责任",
    "解除协议"
])

patient_records = load_ehr("patient_123.json")  # 50 项

diagnosis = llm.diagnose(patient_records)