如何获取 RAG 应用程序以添加引文

本指南回顾了让模型在生成响应时引用源文档的哪些部分的方法。

我们将介绍五种方法：

使用工具调用引用文档 ID;
使用工具调用引用文档 ID 并提供文本片段;
直接提示;
检索后处理（即压缩检索到的上下文以使其更具相关性）;
生成后处理（即发出第二个 LLM 调用，用引文注释生成的答案）。

我们通常建议使用列表中适用于您的用例的第一项。也就是说，如果您的模型支持工具调用，请尝试方法 1 或 2;否则，或者如果这些失败，则向下推进列表。

让我们首先创建一个简单的 RAG 链。首先，我们将使用 WikipediaRetriever 从 Wikipedia 中检索。我们将使用 RAG 教程中的相同 LangGraph 实现。

设置

首先，我们需要安装一些依赖项：

%pip install -qU langchain-community wikipedia

让我们首先选择一个 LLM：

选择聊天模式：

pip install -qU "langchain[openai]"

import getpass
import os

if not os.environ.get("OPENAI_API_KEY"):
  os.environ["OPENAI_API_KEY"] = getpass.getpass("Enter API key for OpenAI: ")

from langchain.chat_models import init_chat_model

llm = init_chat_model("gpt-4o-mini", model_provider="openai")

我们现在可以加载一个检索器并构建我们的提示符：

from langchain_community.retrievers import WikipediaRetriever
from langchain_core.prompts import ChatPromptTemplate

system_prompt = (
    "You're a helpful AI assistant. Given a user question "
    "and some Wikipedia article snippets, answer the user "
    "question. If none of the articles answer the question, "
    "just say you don't know."
    "\n\nHere are the Wikipedia articles: "
    "{context}"
)

retriever = WikipediaRetriever(top_k_results=6, doc_content_chars_max=2000)
prompt = ChatPromptTemplate.from_messages(
    [
        ("system", system_prompt),
        ("human", "{question}"),
    ]
)
prompt.pretty_print()

API 参考：WikipediaRetriever | 聊天提示模板

================================[1m System Message [0m================================

You're a helpful AI assistant. Given a user question and some Wikipedia article snippets, answer the user question. If none of the articles answer the question, just say you don't know.

Here are the Wikipedia articles: [33;1m[1;3m{context}[0m

================================[1m Human Message [0m=================================

[33;1m[1;3m{question}[0m

现在我们已经有了 model、retriever 和 prompt，让我们将它们全部链接在一起。按照向 RAG 应用程序添加引文的操作指南，我们将使我们的链同时返回答案和检索到的文档。这使用与 RAG 教程中相同的 LangGraph 实现。

from langchain_core.documents import Document
from langgraph.graph import START, StateGraph
from typing_extensions import List, TypedDict


# Define state for application
class State(TypedDict):
    question: str
    context: List[Document]
    answer: str


# Define application steps
def retrieve(state: State):
    retrieved_docs = retriever.invoke(state["question"])
    return {"context": retrieved_docs}


def generate(state: State):
    docs_content = "\n\n".join(doc.page_content for doc in state["context"])
    messages = prompt.invoke({"question": state["question"], "context": docs_content})
    response = llm.invoke(messages)
    return {"answer": response.content}


# Compile application and test
graph_builder = StateGraph(State).add_sequence([retrieve, generate])
graph_builder.add_edge(START, "retrieve")
graph = graph_builder.compile()

API 参考：文档 | StateGraph

from IPython.display import Image, display

display(Image(graph.get_graph().draw_mermaid_png()))

result = graph.invoke({"question": "How fast are cheetahs?"})

sources = [doc.metadata["source"] for doc in result["context"]]
print(f"Sources: {sources}\n\n")
print(f'Answer: {result["answer"]}')

Sources: ['https://en.wikipedia.org/wiki/Cheetah', 'https://en.wikipedia.org/wiki/Southeast_African_cheetah', 'https://en.wikipedia.org/wiki/Footspeed', 'https://en.wikipedia.org/wiki/Fastest_animals', 'https://en.wikipedia.org/wiki/Pursuit_predation', 'https://en.wikipedia.org/wiki/Gepard-class_fast_attack_craft']

Answer: Cheetahs are capable of running at speeds between 93 to 104 km/h (58 to 65 mph).

查看 LangSmith 跟踪。

工具调用

如果您选择的 LLM 实现了工具调用功能，您可以使用它来使模型在生成答案时指定它引用了哪些提供的文档。LangChain 工具调用模型实现了.with_structured_output方法，该方法将强制生成遵循所需的 schema（请参阅此处的详细信息）。

引用文档

要使用标识符引用文档，我们将标识符格式化到 Prompt 中，然后使用.with_structured_output强制 LLM 在其输出中引用这些标识符。

首先，我们为输出定义一个架构。这.with_structured_output支持多种格式，包括 JSON 架构和 Pydantic。这里我们将使用 Pydantic：

from pydantic import BaseModel, Field


class CitedAnswer(BaseModel):
    """Answer the user question based only on the given sources, and cite the sources used."""

    answer: str = Field(
        ...,
        description="The answer to the user question, which is based only on the given sources.",
    )
    citations: List[int] = Field(
        ...,
        description="The integer IDs of the SPECIFIC sources which justify the answer.",
    )

让我们看看当我们传入函数和用户输入时，模型输出是什么样的：

structured_llm = llm.with_structured_output(CitedAnswer)

example_q = """What Brian's height?

Source: 1
Information: Suzy is 6'2"

Source: 2
Information: Jeremiah is blonde

Source: 3
Information: Brian is 3 inches shorter than Suzy"""
result = structured_llm.invoke(example_q)

result

CitedAnswer(answer='Brian is 5\'11".', citations=[1, 3])

或者作为 dict：

result.dict()

{'answer': 'Brian is 5\'11".', 'citations': [1, 3]}

现在我们将源标识符构建到 prompt 中，以便使用我们的链进行复制。我们将进行三项更改：

更新提示以包含源标识符;
使用structured_llm（即llm.with_structured_output(CitedAnswer));
在输出中返回 Pydantic 对象。

def format_docs_with_id(docs: List[Document]) -> str:
    formatted = [
        f"Source ID: {i}\nArticle Title: {doc.metadata['title']}\nArticle Snippet: {doc.page_content}"
        for i, doc in enumerate(docs)
    ]
    return "\n\n" + "\n\n".join(formatted)


class State(TypedDict):
    question: str
    context: List[Document]
    answer: CitedAnswer


def generate(state: State):
    formatted_docs = format_docs_with_id(state["context"])
    messages = prompt.invoke({"question": state["question"], "context": formatted_docs})
    structured_llm = llm.with_structured_output(CitedAnswer)
    response = structured_llm.invoke(messages)
    return {"answer": response}


graph_builder = StateGraph(State).add_sequence([retrieve, generate])
graph_builder.add_edge(START, "retrieve")
graph = graph_builder.compile()

result = graph.invoke({"question": "How fast are cheetahs?"})

result["answer"]

CitedAnswer(answer='Cheetahs are capable of running at speeds between 93 to 104 km/h (58 to 65 mph).', citations=[0, 3])

我们可以检查索引 0 处的文档，模型引用了该索引：

print(result["context"][0])

page_content='The cheetah (Acinonyx jubatus) is a large cat and the fastest land animal. It has a tawny to creamy white or pale buff fur that is marked with evenly spaced, solid black spots. The head is small and rounded, with a short snout and black tear-like facial streaks. It reaches 67–94 cm (26–37 in) at the shoulder, and the head-and-body length is between 1.1 and 1.5 m (3 ft 7 in and 4 ft 11 in). Adults weigh between 21 and 72 kg (46 and 159 lb). The cheetah is capable of running at 93 to 104 km/h (58 to 65 mph); it has evolved specialized adaptations for speed, including a light build, long thin legs and a long tail.
The cheetah was first described in the late 18th century. Four subspecies are recognised today that are native to Africa and central Iran. An African subspecies was introduced to India in 2022. It is now distributed mainly in small, fragmented populations in northwestern, eastern and southern Africa and central Iran. It lives in a variety of habitats such as savannahs in the Serengeti, arid mountain ranges in the Sahara, and hilly desert terrain.
The cheetah lives in three main social groups: females and their cubs, male "coalitions", and solitary males. While females lead a nomadic life searching for prey in large home ranges, males are more sedentary and instead establish much smaller territories in areas with plentiful prey and access to females. The cheetah is active during the day, with peaks during dawn and dusk. It feeds on small- to medium-sized prey, mostly weighing under 40 kg (88 lb), and prefers medium-sized ungulates such as impala, springbok and Thomson's gazelles. The cheetah typically stalks its prey within 60–100 m (200–330 ft) before charging towards it, trips it during the chase and bites its throat to suffocate it to death. It breeds throughout the year. After a gestation of nearly three months, females give birth to a litter of three or four cubs. Cheetah cubs are highly vulnerable to predation by other large carnivores. They are weaned a' metadata={'title': 'Cheetah', 'summary': 'The cheetah (Acinonyx jubatus) is a large cat and the fastest land animal. It has a tawny to creamy white or pale buff fur that is marked with evenly spaced, solid black spots. The head is small and rounded, with a short snout and black tear-like facial streaks. It reaches 67–94 cm (26–37 in) at the shoulder, and the head-and-body length is between 1.1 and 1.5 m (3 ft 7 in and 4 ft 11 in). Adults weigh between 21 and 72 kg (46 and 159 lb). The cheetah is capable of running at 93 to 104 km/h (58 to 65 mph); it has evolved specialized adaptations for speed, including a light build, long thin legs and a long tail.\nThe cheetah was first described in the late 18th century. Four subspecies are recognised today that are native to Africa and central Iran. An African subspecies was introduced to India in 2022. It is now distributed mainly in small, fragmented populations in northwestern, eastern and southern Africa and central Iran. It lives in a variety of habitats such as savannahs in the Serengeti, arid mountain ranges in the Sahara, and hilly desert terrain.\nThe cheetah lives in three main social groups: females and their cubs, male "coalitions", and solitary males. While females lead a nomadic life searching for prey in large home ranges, males are more sedentary and instead establish much smaller territories in areas with plentiful prey and access to females. The cheetah is active during the day, with peaks during dawn and dusk. It feeds on small- to medium-sized prey, mostly weighing under 40 kg (88 lb), and prefers medium-sized ungulates such as impala, springbok and Thomson\'s gazelles. The cheetah typically stalks its prey within 60–100 m (200–330 ft) before charging towards it, trips it during the chase and bites its throat to suffocate it to death. It breeds throughout the year. After a gestation of nearly three months, females give birth to a litter of three or four cubs. Cheetah cubs are highly vulnerable to predation by other large carnivores. They are weaned at around four months and are independent by around 20 months of age.\nThe cheetah is threatened by habitat loss, conflict with humans, poaching and high susceptibility to diseases. The global cheetah population was estimated in 2021 at 6,517; it is listed as Vulnerable on the IUCN Red List. It has been widely depicted in art, literature, advertising, and animation. It was tamed in ancient Egypt and trained for hunting ungulates in the Arabian Peninsula and India. It has been kept in zoos since the early 19th century.', 'source': 'https://en.wikipedia.org/wiki/Cheetah'}

LangSmith 跟踪：https://smith.langchain.com/public/6f34d136-451d-4625-90c8-2d8decebc21a/r

引用片段

要返回文本 spans（可能除了源标识符之外），我们可以使用相同的方法。唯一的变化是构建一个更复杂的输出模式，这里使用 Pydantic，其中包括一个 “quote” 和一个源标识符。

旁白：请注意，如果我们拆分文档，以便有许多文档只有一两个句子，而不是一些长文档，那么引用文档就大致等同于引用片段，并且对模型来说可能更容易，因为模型只需要为每个片段返回一个标识符，而不是实际的文本。可能值得尝试这两种方法并进行评估。

class Citation(BaseModel):
    source_id: int = Field(
        ...,
        description="The integer ID of a SPECIFIC source which justifies the answer.",
    )
    quote: str = Field(
        ...,
        description="The VERBATIM quote from the specified source that justifies the answer.",
    )


class QuotedAnswer(BaseModel):
    """Answer the user question based only on the given sources, and cite the sources used."""

    answer: str = Field(
        ...,
        description="The answer to the user question, which is based only on the given sources.",
    )
    citations: List[Citation] = Field(
        ..., description="Citations from the given sources that justify the answer."
    )

class State(TypedDict):
    question: str
    context: List[Document]
    answer: QuotedAnswer


def generate(state: State):
    formatted_docs = format_docs_with_id(state["context"])
    messages = prompt.invoke({"question": state["question"], "context": formatted_docs})
    structured_llm = llm.with_structured_output(QuotedAnswer)
    response = structured_llm.invoke(messages)
    return {"answer": response}


graph_builder = StateGraph(State).add_sequence([retrieve, generate])
graph_builder.add_edge(START, "retrieve")
graph = graph_builder.compile()

在这里，我们看到模型从源 0 中提取了相关的文本片段：

result = graph.invoke({"question": "How fast are cheetahs?"})

result["answer"]

QuotedAnswer(answer='Cheetahs are capable of running at speeds of 93 to 104 km/h (58 to 65 mph).', citations=[Citation(source_id=0, quote='The cheetah is capable of running at 93 to 104 km/h (58 to 65 mph); it has evolved specialized adaptations for speed.')])

LangSmith 跟踪：https://smith.langchain.com/public/e16dc72f-4261-4f25-a9a7-906238737283/r

直接提示

某些模型不支持函数调用。我们可以通过直接提示获得类似的结果。让我们尝试指示模型为其输出生成结构化 XML：

xml_system = """You're a helpful AI assistant. Given a user question and some Wikipedia article snippets, \
answer the user question and provide citations. If none of the articles answer the question, just say you don't know.

Remember, you must return both an answer and citations. A citation consists of a VERBATIM quote that \
justifies the answer and the ID of the quote article. Return a citation for every quote across all articles \
that justify the answer. Use the following format for your final output:

<cited_answer>
    <answer></answer>
    <citations>
        <citation><source_id></source_id><quote></quote></citation>
        <citation><source_id></source_id><quote></quote></citation>
        ...
    </citations>
</cited_answer>

Here are the Wikipedia articles:{context}"""
xml_prompt = ChatPromptTemplate.from_messages(
    [("system", xml_system), ("human", "{question}")]
)

现在，我们对我们的链进行了类似的小更新：

我们更新了格式化函数以将检索到的上下文包装在 XML 标签中;
我们不使用.with_structured_output（例如，因为它对于模型不存在）;
我们使用 XMLOutputParser 将答案解析为字典。

from langchain_core.output_parsers import XMLOutputParser


def format_docs_xml(docs: List[Document]) -> str:
    formatted = []
    for i, doc in enumerate(docs):
        doc_str = f"""\
    <source id=\"{i}\">
        <title>{doc.metadata['title']}</title>
        <article_snippet>{doc.page_content}</article_snippet>
    </source>"""
        formatted.append(doc_str)
    return "\n\n<sources>" + "\n".join(formatted) + "</sources>"


class State(TypedDict):
    question: str
    context: List[Document]
    answer: dict


def generate(state: State):
    formatted_docs = format_docs_xml(state["context"])
    messages = xml_prompt.invoke(
        {"question": state["question"], "context": formatted_docs}
    )
    response = llm.invoke(messages)
    parsed_response = XMLOutputParser().invoke(response)
    return {"answer": parsed_response}


graph_builder = StateGraph(State).add_sequence([retrieve, generate])
graph_builder.add_edge(START, "retrieve")
graph = graph_builder.compile()

API 参考：XMLOutputParser

请注意，引文再次被结构化为答案：

result = graph.invoke({"question": "How fast are cheetahs?"})

result["answer"]

{'cited_answer': [{'answer': 'Cheetahs can run at speeds of 93 to 104 km/h (58 to 65 mph).'},
  {'citations': [{'citation': [{'source_id': '0'},
      {'quote': 'The cheetah is capable of running at 93 to 104 km/h (58 to 65 mph);'}]},
    {'citation': [{'source_id': '3'},
      {'quote': 'The fastest land animal is the cheetah.'}]}]}]}

LangSmith 跟踪：https://smith.langchain.com/public/0c45f847-c640-4b9a-a5fa-63559e413527/r

检索后处理

另一种方法是对检索到的文档进行后处理以压缩内容，这样源内容就已经足够小，我们不需要模型引用特定的源或跨度。例如，我们可以将每个文档分解成一两个句子，嵌入这些文档并只保留最相关的文档。LangChain 为此提供了一些内置组件。在这里，我们将使用 RecursiveCharacterTextSplitter，它通过对分隔符子字符串进行拆分来创建指定大小的块，以及 EmbeddingsFilter，它只保留具有最相关嵌入的文本。

这种方法有效地更新了我们的retrieve步骤压缩文档。让我们首先选择一个嵌入模型：

选择嵌入模型：

pip install -qU langchain-openai

import getpass
import os

if not os.environ.get("OPENAI_API_KEY"):
  os.environ["OPENAI_API_KEY"] = getpass.getpass("Enter API key for OpenAI: ")

from langchain_openai import OpenAIEmbeddings

embeddings = OpenAIEmbeddings(model="text-embedding-3-large")

我们现在可以重写retrieve步：

from langchain.retrievers.document_compressors import EmbeddingsFilter
from langchain_core.runnables import RunnableParallel
from langchain_text_splitters import RecursiveCharacterTextSplitter

splitter = RecursiveCharacterTextSplitter(
    chunk_size=400,
    chunk_overlap=0,
    separators=["\n\n", "\n", ".", " "],
    keep_separator=False,
)
compressor = EmbeddingsFilter(embeddings=embeddings, k=10)


class State(TypedDict):
    question: str
    context: List[Document]
    answer: str


def retrieve(state: State):
    retrieved_docs = retriever.invoke(state["question"])
    split_docs = splitter.split_documents(retrieved_docs)
    stateful_docs = compressor.compress_documents(split_docs, state["question"])
    return {"context": stateful_docs}

API 参考：EmbeddingsFilter | RunnableParallel | 递归角色文本拆分器

让我们测试一下：

retrieval_result = retrieve({"question": "How fast are cheetahs?"})

for doc in retrieval_result["context"]:
    print(f"{doc.page_content}\n\n")

Adults weigh between 21 and 72 kg (46 and 159 lb). The cheetah is capable of running at 93 to 104 km/h (58 to 65 mph); it has evolved specialized adaptations for speed, including a light build, long thin legs and a long tail

The cheetah (Acinonyx jubatus) is a large cat and the fastest land animal. It has a tawny to creamy white or pale buff fur that is marked with evenly spaced, solid black spots. The head is small and rounded, with a short snout and black tear-like facial streaks. It reaches 67–94 cm (26–37 in) at the shoulder, and the head-and-body length is between 1.1 and 1.5 m (3 ft 7 in and 4 ft 11 in)

2 mph), or 171 body lengths per second. The cheetah, the fastest land mammal, scores at only 16 body lengths per second

It feeds on small- to medium-sized prey, mostly weighing under 40 kg (88 lb), and prefers medium-sized ungulates such as impala, springbok and Thomson's gazelles. The cheetah typically stalks its prey within 60–100 m (200–330 ft) before charging towards it, trips it during the chase and bites its throat to suffocate it to death. It breeds throughout the year

The cheetah was first described in the late 18th century. Four subspecies are recognised today that are native to Africa and central Iran. An African subspecies was introduced to India in 2022. It is now distributed mainly in small, fragmented populations in northwestern, eastern and southern Africa and central Iran

The cheetah lives in three main social groups: females and their cubs, male "coalitions", and solitary males. While females lead a nomadic life searching for prey in large home ranges, males are more sedentary and instead establish much smaller territories in areas with plentiful prey and access to females. The cheetah is active during the day, with peaks during dawn and dusk

The Southeast African cheetah (Acinonyx jubatus jubatus) is the nominate cheetah subspecies native to East and Southern Africa. The Southern African cheetah lives mainly in the lowland areas and deserts of the Kalahari, the savannahs of Okavango Delta, and the grasslands of the Transvaal region in South Africa. In Namibia, cheetahs are mostly found in farmlands

Subpopulations have been called "South African cheetah" and "Namibian cheetah."

In India, four cheetahs of the subspecies are living in Kuno National Park in Madhya Pradesh after having been introduced there

Acinonyx jubatus velox proposed in 1913 by Edmund Heller on basis of a cheetah that was shot by Kermit Roosevelt in June 1909 in the Kenyan highlands.
Acinonyx rex proposed in 1927 by Reginald Innes Pocock on basis of a specimen from the Umvukwe Range in Rhodesia.

接下来，我们像以前一样将其组装到我们的链中：

# This step is unchanged from our original RAG implementation
def generate(state: State):
    docs_content = "\n\n".join(doc.page_content for doc in state["context"])
    messages = prompt.invoke({"question": state["question"], "context": docs_content})
    response = llm.invoke(messages)
    return {"answer": response.content}


graph_builder = StateGraph(State).add_sequence([retrieve, generate])
graph_builder.add_edge(START, "retrieve")
graph = graph_builder.compile()

result = graph.invoke({"question": "How fast are cheetahs?"})

print(result["answer"])

Cheetahs are capable of running at speeds between 93 to 104 km/h (58 to 65 mph). They are known as the fastest land animals.

请注意，文档内容现在已压缩，尽管文档对象在其元数据的“summary”键中保留了原始内容。这些摘要不会传递给模型;只有 Condensed Content 才是。

result["context"][0].page_content  # passed to model

'Adults weigh between 21 and 72 kg (46 and 159 lb). The cheetah is capable of running at 93 to 104 km/h (58 to 65 mph); it has evolved specialized adaptations for speed, including a light build, long thin legs and a long tail'

result["context"][0].metadata["summary"]  # original document  # original document

'The cheetah (Acinonyx jubatus) is a large cat and the fastest land animal. It has a tawny to creamy white or pale buff fur that is marked with evenly spaced, solid black spots. The head is small and rounded, with a short snout and black tear-like facial streaks. It reaches 67–94 cm (26–37 in) at the shoulder, and the head-and-body length is between 1.1 and 1.5 m (3 ft 7 in and 4 ft 11 in). Adults weigh between 21 and 72 kg (46 and 159 lb). The cheetah is capable of running at 93 to 104 km/h (58 to 65 mph); it has evolved specialized adaptations for speed, including a light build, long thin legs and a long tail.\nThe cheetah was first described in the late 18th century. Four subspecies are recognised today that are native to Africa and central Iran. An African subspecies was introduced to India in 2022. It is now distributed mainly in small, fragmented populations in northwestern, eastern and southern Africa and central Iran. It lives in a variety of habitats such as savannahs in the Serengeti, arid mountain ranges in the Sahara, and hilly desert terrain.\nThe cheetah lives in three main social groups: females and their cubs, male "coalitions", and solitary males. While females lead a nomadic life searching for prey in large home ranges, males are more sedentary and instead establish much smaller territories in areas with plentiful prey and access to females. The cheetah is active during the day, with peaks during dawn and dusk. It feeds on small- to medium-sized prey, mostly weighing under 40 kg (88 lb), and prefers medium-sized ungulates such as impala, springbok and Thomson\'s gazelles. The cheetah typically stalks its prey within 60–100 m (200–330 ft) before charging towards it, trips it during the chase and bites its throat to suffocate it to death. It breeds throughout the year. After a gestation of nearly three months, females give birth to a litter of three or four cubs. Cheetah cubs are highly vulnerable to predation by other large carnivores. They are weaned at around four months and are independent by around 20 months of age.\nThe cheetah is threatened by habitat loss, conflict with humans, poaching and high susceptibility to diseases. The global cheetah population was estimated in 2021 at 6,517; it is listed as Vulnerable on the IUCN Red List. It has been widely depicted in art, literature, advertising, and animation. It was tamed in ancient Egypt and trained for hunting ungulates in the Arabian Peninsula and India. It has been kept in zoos since the early 19th century.'

LangSmith 跟踪：https://smith.langchain.com/public/21b0dc15-d70a-4293-9402-9c70f9178e66/r

生成后处理

另一种方法是对模型生成进行后处理。在此示例中，我们首先只生成一个答案，然后我们要求模型用引文注释自己的答案。这种方法的缺点当然是速度较慢且成本较高，因为需要进行两次模型调用。

让我们将其应用于我们的初始链。如果需要，我们可以通过应用程序中的第三步来实现这一点。

class Citation(BaseModel):
    source_id: int = Field(
        ...,
        description="The integer ID of a SPECIFIC source which justifies the answer.",
    )
    quote: str = Field(
        ...,
        description="The VERBATIM quote from the specified source that justifies the answer.",
    )


class AnnotatedAnswer(BaseModel):
    """Annotate the answer to the user question with quote citations that justify the answer."""

    citations: List[Citation] = Field(
        ..., description="Citations from the given sources that justify the answer."
    )


structured_llm = llm.with_structured_output(AnnotatedAnswer)

class State(TypedDict):
    question: str
    context: List[Document]
    answer: str
    annotations: AnnotatedAnswer


def retrieve(state: State):
    retrieved_docs = retriever.invoke(state["question"])
    return {"context": retrieved_docs}


def generate(state: State):
    docs_content = "\n\n".join(doc.page_content for doc in state["context"])
    messages = prompt.invoke({"question": state["question"], "context": docs_content})
    response = llm.invoke(messages)
    return {"answer": response.content}


def annotate(state: State):
    formatted_docs = format_docs_with_id(state["context"])
    messages = [
        ("system", system_prompt.format(context=formatted_docs)),
        ("human", state["question"]),
        ("ai", state["answer"]),
        ("human", "Annotate your answer with citations."),
    ]
    response = structured_llm.invoke(messages)
    return {"annotations": response}


graph_builder = StateGraph(State).add_sequence([retrieve, generate, annotate])
graph_builder.add_edge(START, "retrieve")
graph = graph_builder.compile()

display(Image(graph.get_graph().draw_mermaid_png()))

result = graph.invoke({"question": "How fast are cheetahs?"})

print(result["answer"])

Cheetahs are capable of running at speeds between 93 to 104 km/h (58 to 65 mph).

result["annotations"]

AnnotatedAnswer(citations=[Citation(source_id=0, quote='The cheetah is capable of running at 93 to 104 km/h (58 to 65 mph)')])

LangSmith 跟踪： https://smith.langchain.com/public/b8257417-573b-47c4-a750-74e542035f19/r