Pydantic Extractor¶

这里我们测试一下 PydanticProgramExtractor 的能力 - 能够使用大型语言模型（无论是标准的文本补全 LLM 还是函数调用 LLM）提取出一个完整的 Pydantic 对象。

与使用“单个”元数据提取器相比，其优势在于我们可以在一次 LLM 调用中提取多个实体。

设置¶

In [ ]

%pip install llama-index-readers-web
%pip install llama-index-program-openai

%pip install llama-index-readers-web %pip install llama-index-program-openai

In [ ]

import nest_asyncio

nest_asyncio.apply()

import os
import openai

import nest_asyncio nest_asyncio.apply() import os import openai

In [ ]

os.environ["OPENAI_API_KEY"] = "YOUR_API_KEY"
openai.api_key = os.getenv("OPENAI_API_KEY")

os.environ["OPENAI_API_KEY"] = "YOUR_API_KEY" openai.api_key = os.getenv("OPENAI_API_KEY")

设置 Pydantic 模型¶

这里我们定义了一个想要提取的基本结构化模式。它包含：

• entities：文本块中的唯一实体
• summary：文本块的简洁摘要
• contains_number：块是否包含数字

这显然是一个示例模式。我们鼓励您在想要提取的元数据类型上发挥创意！

In [ ]

from pydantic import BaseModel, Field
from typing import List

from pydantic import BaseModel, Field from typing import List

In [ ]

class NodeMetadata(BaseModel):
    """Node metadata."""

    entities: List[str] = Field(
        ..., description="Unique entities in this text chunk."
    )
    summary: str = Field(
        ..., description="A concise summary of this text chunk."
    )
    contains_number: bool = Field(
        ...,
        description=(
            "Whether the text chunk contains any numbers (ints, floats, etc.)"
        ),
    )

class NodeMetadata(BaseModel): """节点元数据""" entities: List[str] = Field( ..., description="此文本块中的唯一实体。" ) summary: str = Field( ..., description="此文本块的简洁摘要。" ) contains_number: bool = Field( ..., description=( "此文本块是否包含任何数字（整数、浮点数等）" ), )

设置提取器¶

这里我们设置元数据提取器。请注意，我们提供了提示模板，以便了解其内部工作原理。

In [ ]

from llama_index.program.openai import OpenAIPydanticProgram
from llama_index.core.extractors import PydanticProgramExtractor

EXTRACT_TEMPLATE_STR = """\
Here is the content of the section:
----------------
{context_str}
----------------
Given the contextual information, extract out a {class_name} object.\
"""

openai_program = OpenAIPydanticProgram.from_defaults(
    output_cls=NodeMetadata,
    prompt_template_str="{input}",
    # extract_template_str=EXTRACT_TEMPLATE_STR
)

program_extractor = PydanticProgramExtractor(
    program=openai_program, input_key="input", show_progress=True
)

from llama_index.program.openai import OpenAIPydanticProgram from llama_index.core.extractors import PydanticProgramExtractor EXTRACT_TEMPLATE_STR = """\ 这是本节的内容： ---------------- {context_str} ---------------- 根据上下文信息，提取出一个 {class_name} 对象。\ """ openai_program = OpenAIPydanticProgram.from_defaults( output_cls=NodeMetadata, prompt_template_str="{input}", # extract_template_str=EXTRACT_TEMPLATE_STR ) program_extractor = PydanticProgramExtractor( program=openai_program, input_key="input", show_progress=True )

载入数据¶

我们使用 LlamaHub 的 SimpleWebPageReader 载入 Eugene 的文章（https://eugeneyan.com/writing/llm-patterns/）。

In [ ]

# load in blog

from llama_index.readers.web import SimpleWebPageReader
from llama_index.core.node_parser import SentenceSplitter

reader = SimpleWebPageReader(html_to_text=True)
docs = reader.load_data(urls=["https://eugeneyan.com/writing/llm-patterns/"])

# 载入博客 from llama_index.readers.web import SimpleWebPageReader from llama_index.core.node_parser import SentenceSplitter reader = SimpleWebPageReader(html_to_text=True) docs = reader.load_data(urls=["https://eugeneyan.com/writing/llm-patterns/"])

In [ ]

from llama_index.core.ingestion import IngestionPipeline

node_parser = SentenceSplitter(chunk_size=1024)

pipeline = IngestionPipeline(transformations=[node_parser, program_extractor])

orig_nodes = pipeline.run(documents=docs)

from llama_index.core.ingestion import IngestionPipeline node_parser = SentenceSplitter(chunk_size=1024) pipeline = IngestionPipeline(transformations=[node_parser, program_extractor]) orig_nodes = pipeline.run(documents=docs)

In [ ]

orig_nodes

orig_nodes

提取元数据¶

既然我们已经设置好了元数据提取器和数据，就可以开始提取一些元数据了！

我们看到 Pydantic 特征提取器能够在一次 LLM 调用中从给定的文本块中提取*所有*元数据。

In [ ]

sample_entry = program_extractor.extract(orig_nodes[0:1])[0]

sample_entry = program_extractor.extract(orig_nodes[0:1])[0]

Extracting Pydantic object:   0%|          | 0/1 [00:00<?, ?it/s]

In [ ]

display(sample_entry)

display(sample_entry)

{'entities': ['eugeneyan', 'HackerNews', 'Karpathy'],
 'summary': 'This section discusses practical patterns for integrating large language models (LLMs) into systems & products. It introduces seven key patterns and provides information on evaluations and benchmarks in the field of language modeling.',
 'contains_number': True}

In [ ]

new_nodes = program_extractor.process_nodes(orig_nodes)

new_nodes = program_extractor.process_nodes(orig_nodes)

Extracting Pydantic object:   0%|          | 0/29 [00:00<?, ?it/s]

In [ ]

display(new_nodes[5:7])

display(new_nodes[5:7])