本电子书开源,欢迎 star 🌟,关注《LLM 应用开发实践笔记》

我的新书《LangChain编程从入门到实践》 已经开售!推荐正在学习AI应用开发的朋友购买阅读! LangChain编程从入门到实践

LangChain模块之 Chains

链定义为对组件的一系列调用,也可以包括其他链,这种在链中将组件组合在一起的想法很简单但功能强大,极大地简化了复杂应用程序的实现并使其更加模块化,这反过来又使调试、维护和改进应用程序变得更加容易。 Chain基类是所有chain对象的基本入口,与用户程序交互,处理用户的输入,准备其他模块的输入,提供内存能力,chain的回调能力,其他所有的 Chain 类都继承自这个基类,并根据需要实现特定的功能。

class Chain(BaseModel, ABC):
    memory: BaseMemory
    callbacks: Callbacks

    def __call__(
        self,
        inputs: Any,
        return_only_outputs: bool = False,
        callbacks: Callbacks = None,
    ) -> Dict[str, Any]:
        ...

实现自定义链

from __future__ import annotations

from typing import Any, Dict, List, Optional

from pydantic import Extra

from langchain.base_language import BaseLanguageModel
from langchain.callbacks.manager import (
    AsyncCallbackManagerForChainRun,
    CallbackManagerForChainRun,
)
from langchain.chains.base import Chain
from langchain.prompts.base import BasePromptTemplate


class MyCustomChain(Chain):
    prompt: BasePromptTemplate
    llm: BaseLanguageModel
    output_key: str = "text"

    class Config:
        extra = Extra.forbid
        arbitrary_types_allowed = True

    @property
    def input_keys(self) -> List[str]:
        """pompt中的动态变量
        """
        return self.prompt.input_variables

    @property
    def output_keys(self) -> List[str]:
        """允许直接输出的动态变量.
        """
        return [self.output_key]
    # 同步调用
    def _call(
        self,
        inputs: Dict[str, Any],
        run_manager: Optional[CallbackManagerForChainRun] = None,
    ) -> Dict[str, str]:
        # 下面是一个自定义逻辑实现
        prompt_value = self.prompt.format_prompt(**inputs)
        # 调用一个语言模型或另一个链时,传递一个回调处理。这样内部运行可以通过这个回调(进行逻辑处理)。
        response = self.llm.generate_prompt(
            [prompt_value], callbacks=run_manager.get_child() if run_manager else None
        )
        # 回调出发时的日志输出
        if run_manager:
            run_manager.on_text("Log something about this run")

        return {self.output_key: response.generations[0][0].text}
    # 异步调用
    async def _acall(
        self,
        inputs: Dict[str, Any],
        run_manager: Optional[AsyncCallbackManagerForChainRun] = None,
    ) -> Dict[str, str]:
        prompt_value = self.prompt.format_prompt(**inputs)
        response = await self.llm.agenerate_prompt(
            [prompt_value], callbacks=run_manager.get_child() if run_manager else None
        )
        if run_manager:
            await run_manager.on_text("Log something about this run")

        return {self.output_key: response.generations[0][0].text}

    @property
    def _chain_type(self) -> str:
        return "my_custom_chain"

继承Chain的子类主要有两种类型:

通用工具 chain: 控制chain的调用顺序, 是否调用,他们可以用来合并构造其他的chain。 专门用途 chain: 和通用chain比较来说,他们承担了具体的某项任务,可以和通用的chain组合起来使用,也可以直接使用。有些 Chain 类可能用于处理文本数据,有些可能用于处理图像数据,有些可能用于处理音频数据等。

从 LangChainHub 加载链

LangChainHub 托管了一些高质量Prompt、Agent和Chain,可以直接在langchain中使用。

def test_mathchain():
    from langchain.chains import load_chain
    chain = load_chain("lc://chains/llm-math/chain.json")
    """
    > Entering new  chain...
    2+2等于几Answer: 4
    > Finished chain.
    Answer: 4
    """
    print(chain.run("2+2等于几"))

运行 LLM 链的五种方式

from langchain import PromptTemplate, OpenAI, LLMChain

prompt_template = "给做 {product} 的公司起一个名字?"

llm = OpenAI(temperature=0)
llm_chain = LLMChain(
    llm=llm,
    prompt=PromptTemplate.from_template(prompt_template)
)
print(llm_chain("儿童玩具"))
print(llm_chain.run("儿童玩具"))
llm_chain.apply([{"product":"儿童玩具"}])
llm_chain.generate([{"product":"儿童玩具"}])
llm_chain.predict(product="儿童玩具")

通用工具chain

  1. MultiPromptChain:可以动态选择与给定问题最相关的提示,然后使用该提示回答问题。
  2. EmbeddingRouterChain:使用嵌入和相似性动态选择下一个链。
  3. LLMRouterChain:使用 LLM 来确定动态选择下一个链。
  4. SimpleSequentialChain/SequentialChain:将多个链按照顺序组成处理流水线,SimpleMemory支持在多个链之间传递上下文
  5. TransformChain:一个自定义方法做动态转换的链 ```python def transform_func(inputs: dict) -> dict:
     text = inputs["text"]
     shortened_text = "\n\n".join(text.split("\n\n")[:3])
     return {"output_text": shortened_text}
    
transform_chain = TransformChain(
    input_variables=["text"], output_variables=["output_text"], transform=transform_func
)
template = """Summarize this text:

{output_text}

Summary:"""
prompt = PromptTemplate(input_variables=["output_text"], template=template)
llm_chain = LLMChain(llm=OpenAI(), prompt=prompt)
sequential_chain = SimpleSequentialChain(chains=[transform_chain, llm_chain])
```

合并文档的链(专门用途chain)

BaseCombineDocumentsChain 有四种不同的模式

def load_qa_chain(
    llm: BaseLanguageModel,
    chain_type: str = "stuff",
    verbose: Optional[bool] = None,
    callback_manager: Optional[BaseCallbackManager] = None,
    **kwargs: Any,
) -> BaseCombineDocumentsChain:
    """Load question answering chain.

    Args:
        llm: Language Model to use in the chain.
        chain_type: Type of document combining chain to use. Should be one of "stuff",
            "map_reduce", "map_rerank", and "refine".
        verbose: Whether chains should be run in verbose mode or not. Note that this
            applies to all chains that make up the final chain.
        callback_manager: Callback manager to use for the chain.

    Returns:
        A chain to use for question answering.
    """
    loader_mapping: Mapping[str, LoadingCallable] = {
        "stuff": _load_stuff_chain,
        "map_reduce": _load_map_reduce_chain,
        "refine": _load_refine_chain,
        "map_rerank": _load_map_rerank_chain,
    }

StuffDocumentsChain

获取一个文档列表,带入提示上下文,传递给LLM(适合小文档)

def _load_stuff_chain(
    llm: BaseLanguageModel,
    prompt: Optional[BasePromptTemplate] = None,
    document_variable_name: str = "context",
    verbose: Optional[bool] = None,
    callback_manager: Optional[BaseCallbackManager] = None,
    callbacks: Callbacks = None,
    **kwargs: Any,
) -> StuffDocumentsChain:

RefineDocumentsChain

在Studff方式上进一步优化,循环输入文档并迭代更新其答案,以获得最好的最终结果。具体做法是将所有非文档输入、当前文档和最新的中间答案组合传递给LLM。(适合LLM上下文大小不能容纳的小文档)

def _load_refine_chain(
    llm: BaseLanguageModel,
    question_prompt: Optional[BasePromptTemplate] = None,
    refine_prompt: Optional[BasePromptTemplate] = None,
    document_variable_name: str = "context_str",
    initial_response_name: str = "existing_answer",
    refine_llm: Optional[BaseLanguageModel] = None,
    verbose: Optional[bool] = None,
    callback_manager: Optional[BaseCallbackManager] = None,
    callbacks: Callbacks = None,
    **kwargs: Any,
) -> RefineDocumentsChain:

MapReduceDocumentsChain

将LLM链应用于每个单独的文档(Map步骤),将链的输出视为新文档。然后,将所有新文档传递给单独的合并文档链以获得单一输出(Reduce步骤)。在执行Map步骤前也可以对每个单独文档进行压缩或合并映射,以确保它们适合合并文档链;可以将这个步骤递归执行直到满足要求。(适合大规模文档的情况)

def _load_map_reduce_chain(
    llm: BaseLanguageModel,
    question_prompt: Optional[BasePromptTemplate] = None,
    combine_prompt: Optional[BasePromptTemplate] = None,
    combine_document_variable_name: str = "summaries",
    map_reduce_document_variable_name: str = "context",
    collapse_prompt: Optional[BasePromptTemplate] = None,
    reduce_llm: Optional[BaseLanguageModel] = None,
    collapse_llm: Optional[BaseLanguageModel] = None,
    verbose: Optional[bool] = None,
    callback_manager: Optional[BaseCallbackManager] = None,
    callbacks: Callbacks = None,
    **kwargs: Any,
) -> MapReduceDocumentsChain:

MapRerankDocumentsChain

每个文档上运行一个初始提示,再给对应输出给一个分数,返回得分最高的回答。

def _load_map_rerank_chain(
    llm: BaseLanguageModel,
    prompt: BasePromptTemplate = map_rerank_prompt.PROMPT,
    verbose: bool = False,
    document_variable_name: str = "context",
    rank_key: str = "score",
    answer_key: str = "answer",
    callback_manager: Optional[BaseCallbackManager] = None,
    callbacks: Callbacks = None,
    **kwargs: Any,
) -> MapRerankDocumentsChain:

获取领域知识的链(专门用途chain)

APIChain使得可以使用LLMs与API进行交互,以检索相关信息。通过提供与所提供的API文档相关的问题来构建链。 下面是与播客查询相关的

import os
from langchain.llms import OpenAI
from langchain.chains.api import podcast_docs
from langchain.chains import APIChain

listen_api_key = 'xxx'
llm = OpenAI(temperature=0)
headers = {"X-ListenAPI-Key": listen_api_key}
chain = APIChain.from_llm_and_api_docs(llm, podcast_docs.PODCAST_DOCS, headers=headers, verbose=True)
chain.run("搜索关于ChatGPT的节目, 要求超过30分钟,只返回一条")

合并文档的链的高频使用场景举例

对话场景(最广泛)

ConversationalRetrievalChain 对话式检索链的工作原理:将聊天历史记录(显式传入或从提供的内存中检索)和问题合并到一个独立的问题中,然后从检索器查找相关文档,最后将这些文档和问题传递给问答链以返回响应。

def test_converstion():
    from langchain.chains import ConversationalRetrievalChain
    from langchain.memory import ConversationBufferMemory
    loader = TextLoader("./test.txt")
    documents = loader.load()
    text_splitter = CharacterTextSplitter(chunk_size=1000, chunk_overlap=0)
    documents = text_splitter.split_documents(documents)

    embeddings = OpenAIEmbeddings()
    vectorstore = Chroma.from_documents(documents, embeddings)
    memory = ConversationBufferMemory(memory_key="chat_history", return_messages=True)
    qa = ConversationalRetrievalChain.from_llm(OpenAI(temperature=0), vectorstore.as_retriever(), memory=memory)
    query = "这本书包含哪些内容?"
    result = qa({"question": query})
    print(result)
    chat_history = [(query, result["answer"])]
    query = "还有要补充的吗"
    result = qa({"question": query, "chat_history": chat_history})
    print(result["answer"])

基于数据库问答场景

def test_db_chain():
    from langchain import OpenAI, SQLDatabase, SQLDatabaseChain
    db = SQLDatabase.from_uri("sqlite:///../user.db")
    llm = OpenAI(temperature=0, verbose=True)
    db_chain = SQLDatabaseChain.from_llm(llm, db, verbose=True, use_query_checker=True)
    db_chain.run("有多少用户?")

总结场景

def test_summary():
    from langchain.chains.summarize import load_summarize_chain

    text_splitter = CharacterTextSplitter()
    with open("./测试.txt") as f:
        state_of_the_union = f.read()
    texts = text_splitter.split_text(state_of_the_union)
    docs = [Document(page_content=t) for t in texts[:3]]
    chain = load_summarize_chain(OpenAI(temperature=0), chain_type="map_reduce")
    chain.run(docs)

问答场景

def test_qa():
    from langchain.chains.question_answering import load_qa_chain
    loader = TextLoader("./测试.txt")
    documents = loader.load()
    text_splitter = CharacterTextSplitter(chunk_size=1000, chunk_overlap=0)
    texts = text_splitter.split_documents(documents)
    embeddings = OpenAIEmbeddings()
    docsearch = Chroma.from_documents(texts, embeddings)
    qa_chain = load_qa_chain(OpenAI(temperature=0), chain_type="map_reduce")
    qa = RetrievalQA(combine_documents_chain=qa_chain, retriever=docsearch.as_retriever())
    qa.run()

results matching ""

    No results matching ""