Movatterモバイル変換

[0]ホーム
URL:

Skip to main content
OurBuilding Ambient Agents with LangGraph course is now available on LangChain Academy!
Open In ColabOpen on GitHub

Build an Extraction Chain

In this tutorial, we will usetool-calling features ofchat models to extract structured information from unstructured text. We will also demonstrate how to usefew-shot prompting in this context to improve performance.

important

This tutorial requireslangchain-core>=0.3.20 and will only work with models that supporttool calling.

Setup​

Jupyter Notebook​

This and other tutorials are perhaps most conveniently run in aJupyter notebooks. Going through guides in an interactive environment is a great way to better understand them. Seehere for instructions on how to install.

Installation​

To install LangChain run:

  • Pip
  • Conda
pip install --upgrade langchain-core

For more details, see ourInstallation guide.

LangSmith​

Many of the applications you build with LangChain will contain multiple steps with multiple invocations of LLM calls.As these applications get more and more complex, it becomes crucial to be able to inspect what exactly is going on inside your chain or agent.The best way to do this is withLangSmith.

After you sign up at the link above, make sure to set your environment variables to start logging traces:

export LANGSMITH_TRACING="true"
export LANGSMITH_API_KEY="..."

Or, if in a notebook, you can set them with:

import getpass
import os

os.environ["LANGSMITH_TRACING"]="true"
os.environ["LANGSMITH_API_KEY"]= getpass.getpass()

The Schema​

First, we need to describe what information we want to extract from the text.

We'll use Pydantic to define an example schema to extract personal information.

from typingimport Optional

from pydanticimport BaseModel, Field


classPerson(BaseModel):
"""Information about a person."""

# ^ Doc-string for the entity Person.
# This doc-string is sent to the LLM as the description of the schema Person,
# and it can help to improve extraction results.

# Note that:
# 1. Each field is an `optional` -- this allows the model to decline to extract it!
# 2. Each field has a `description` -- this description is used by the LLM.
# Having a good description can help improve extraction results.
    name: Optional[str]= Field(default=None, description="The name of the person")
    hair_color: Optional[str]= Field(
        default=None, description="The color of the person's hair if known"
)
    height_in_meters: Optional[str]= Field(
        default=None, description="Height measured in meters"
)

There are two best practices when defining schema:

  1. Document theattributes and theschema itself: This information is sent to the LLM and is used to improve the quality of information extraction.
  2. Do not force the LLM to make up information! Above we usedOptional for the attributes allowing the LLM to outputNone if it doesn't know the answer.
important

For best performance, document the schema well and make sure the model isn't forced to return results if there's no information to be extracted in the text.

The Extractor​

Let's create an information extractor using the schema we defined above.

from langchain_core.promptsimport ChatPromptTemplate, MessagesPlaceholder

# Define a custom prompt to provide instructions and any additional context.
# 1) You can add examples into the prompt template to improve extraction quality
# 2) Introduce additional parameters to take context into account (e.g., include metadata
#    about the document from which the text was extracted.)
prompt_template= ChatPromptTemplate.from_messages(
[
(
"system",
"You are an expert extraction algorithm. "
"Only extract relevant information from the text. "
"If you do not know the value of an attribute asked to extract, "
"return null for the attribute's value.",
),
# Please see the how-to about improving performance with
# reference examples.
# MessagesPlaceholder('examples'),
("human","{text}"),
]
)

We need to use a model that supports function/tool calling.

Please reviewthe documentation for all models that can be used with this API.

pip install -qU "langchain[google-genai]"
import getpass
import os

ifnot os.environ.get("GOOGLE_API_KEY"):
  os.environ["GOOGLE_API_KEY"]= getpass.getpass("Enter API key for Google Gemini: ")

from langchain.chat_modelsimport init_chat_model

llm= init_chat_model("gemini-2.0-flash", model_provider="google_genai")
structured_llm= llm.with_structured_output(schema=Person)

Let's test it out:

text="Alan Smith is 6 feet tall and has blond hair."
prompt= prompt_template.invoke({"text": text})
structured_llm.invoke(prompt)
Person(name='Alan Smith', hair_color='blond', height_in_meters='1.83')
important

Extraction is Generative 🤯

LLMs are generative models, so they can do some pretty cool things like correctly extract the height of the person in meterseven though it was provided in feet!

We can see the LangSmith tracehere. Note that thechat model portion of the trace reveals the exact sequence of messages sent to the model, tools invoked, and other metadata.

Multiple Entities​

Inmost cases, you should be extracting a list of entities rather than a single entity.

This can be easily achieved using pydantic by nesting models inside one another.

from typingimport List, Optional

from pydanticimport BaseModel, Field


classPerson(BaseModel):
"""Information about a person."""

# ^ Doc-string for the entity Person.
# This doc-string is sent to the LLM as the description of the schema Person,
# and it can help to improve extraction results.

# Note that:
# 1. Each field is an `optional` -- this allows the model to decline to extract it!
# 2. Each field has a `description` -- this description is used by the LLM.
# Having a good description can help improve extraction results.
    name: Optional[str]= Field(default=None, description="The name of the person")
    hair_color: Optional[str]= Field(
        default=None, description="The color of the person's hair if known"
)
    height_in_meters: Optional[str]= Field(
        default=None, description="Height measured in meters"
)


classData(BaseModel):
"""Extracted data about people."""

# Creates a model so that we can extract multiple entities.
    people: List[Person]
important

Extraction results might not be perfect here. Read on to see how to useReference Examples to improve the quality of extraction, and check out our extractionhow-to guides for more detail.

structured_llm= llm.with_structured_output(schema=Data)
text="My name is Jeff, my hair is black and i am 6 feet tall. Anna has the same color hair as me."
prompt= prompt_template.invoke({"text": text})
structured_llm.invoke(prompt)
Data(people=[Person(name='Jeff', hair_color='black', height_in_meters='1.83'), Person(name='Anna', hair_color='black', height_in_meters=None)])
tip

When the schema accommodates the extraction ofmultiple entities, it also allows the model to extractno entities if no relevant informationis in the text by providing an empty list.

This is usually agood thing! It allows specifyingrequired attributes on an entity without necessarily forcing the model to detect this entity.

We can see the LangSmith tracehere.

Reference examples​

The behavior of LLM applications can be steered usingfew-shot prompting. Forchat models, this can take the form of a sequence of pairs of input and response messages demonstrating desired behaviors.

For example, we can convey the meaning of a symbol with alternatinguser andassistantmessages:

messages=[
{"role":"user","content":"2 🦜 2"},
{"role":"assistant","content":"4"},
{"role":"user","content":"2 🦜 3"},
{"role":"assistant","content":"5"},
{"role":"user","content":"3 🦜 4"},
]

response= llm.invoke(messages)
print(response.content)
7

Structured output often usestool calling under-the-hood. This typically involves the generation ofAI messages containing tool calls, as well astool messages containing the results of tool calls. What should a sequence of messages look like in this case?

Differentchat model providers impose different requirements for valid message sequences. Some will accept a (repeating) message sequence of the form:

  • User message
  • AI message with tool call
  • Tool message with result

Others require a final AI message containing some sort of response.

LangChain includes a utility functiontool_example_to_messages that will generate a valid sequence for most model providers. It simplifies the generation of structured few-shot examples by just requiring Pydantic representations of the corresponding tool calls.

Let's try this out. We can convert pairs of input strings and desired Pydantic objects to a sequence of messages that can be provided to a chat model. Under the hood, LangChain will format the tool calls to each provider's required format.

Note: this version oftool_example_to_messages requireslangchain-core>=0.3.20.

from langchain_core.utils.function_callingimport tool_example_to_messages

examples=[
(
"The ocean is vast and blue. It's more than 20,000 feet deep.",
        Data(people=[]),
),
(
"Fiona traveled far from France to Spain.",
        Data(people=[Person(name="Fiona", height_in_meters=None, hair_color=None)]),
),
]


messages=[]

for txt, tool_callin examples:
if tool_call.people:
# This final message is optional for some providers
        ai_response="Detected people."
else:
        ai_response="Detected no people."
    messages.extend(tool_example_to_messages(txt,[tool_call], ai_response=ai_response))

Inspecting the result, we see these two example pairs generated eight messages:

for messagein messages:
    message.pretty_print()
================================ Human Message =================================

The ocean is vast and blue. It's more than 20,000 feet deep.
================================== Ai Message ==================================
Tool Calls:
  Data (d8f2e054-7fb9-417f-b28f-0447a775b2c3)
 Call ID: d8f2e054-7fb9-417f-b28f-0447a775b2c3
  Args:
    people: []
================================= Tool Message =================================

You have correctly called this tool.
================================== Ai Message ==================================

Detected no people.
================================ Human Message =================================

Fiona traveled far from France to Spain.
================================== Ai Message ==================================
Tool Calls:
  Data (0178939e-a4b1-4d2a-a93e-b87f665cdfd6)
 Call ID: 0178939e-a4b1-4d2a-a93e-b87f665cdfd6
  Args:
    people: [{'name': 'Fiona', 'hair_color': None, 'height_in_meters': None}]
================================= Tool Message =================================

You have correctly called this tool.
================================== Ai Message ==================================

Detected people.

Let's compare performance with and without these messages. For example, let's pass a message for which we intend no people to be extracted:

message_no_extraction={
"role":"user",
"content":"The solar system is large, but earth has only 1 moon.",
}

structured_llm= llm.with_structured_output(schema=Data)
structured_llm.invoke([message_no_extraction])
Data(people=[Person(name='Earth', hair_color='None', height_in_meters='0.00')])

In this example, the model is liable to erroneously generate records of people.

Because our few-shot examples contain examples of "negatives", we encourage the model to behave correctly in this case:

structured_llm.invoke(messages+[message_no_extraction])
Data(people=[])
tip

TheLangSmith trace for the run reveals the exact sequence of messages sent to the chat model, tool calls generated, latency, token counts, and other metadata.

Seethis guide for more detail on extraction workflows with reference examples, including how to incorporate prompt templates and customize the generation of example messages.

Next steps​

Now that you understand the basics of extraction with LangChain, you're ready to proceed to the rest of the how-to guides:

  • Add Examples: More detail on usingreference examples to improve performance.
  • Handle Long Text: What should you do if the text does not fit into the context window of the LLM?
  • Use a Parsing Approach: Use a prompt based approach to extract with models that do not supporttool/function calling.
[8]ページ先頭
©2009-2025 Movatter.jp