当前位置：首页 > news >正文

langchain 结构化输出

news 2025/2/26 3:50:26

主要流程

1. 使用 Pydantic 定义结构化输出：

定义 AnswerWithJustification 类，用于描述输出的结构，包含以下字段：

answer：答案内容（字符串类型）。
justification：答案的理由或解释（字符串类型，部分场景为可选）。
这种方法保证了 LLM 的输出符合预期的结构化格式。

2. with_structured_output 方法：

通过调用 with_structured_output 配置 LLM 以生成结构化的输出。
主要参数包括：

strict=True：严格模式，确保输出必须符合定义的结构。
include_raw=True：在返回的结果中同时包含模型的原始响应（未解析的JSON）以及解析后的对象。

3. 不同的输出结构实现方式：

基于 Pydantic：

使用 AnswerWithJustification 类定义结构。
输出严格符合类定义，模型会返回符合 BaseModel 结构的内容。

基于 JSON Schema：

使用一个 oai_schema 字典直接定义输出的 JSON 格式（包括字段名称、描述及类型）。
JSON Schema 更灵活，适用于无需定义 Python 类的场景。

启用 json_mode：

指定输出为纯 JSON 格式，并返回包含字段 answer 和 justification 的 JSON 数据。
模型调用逻辑：

4. 配置一个名为 GLM-4-flash 的模型，使用指定的 API 密钥和自定义 API 地址。
使用 invoke 方法对问题生成答案，返回值包括：

raw：LLM 的原始 JSON 格式响应。
parsed：基于结构化输出模式解析后的对象（如符合 Pydantic 定义的实例）。
parsing_error：解析错误信息（若有）。

具体代码

1. strict=True

from typing import Optionalfrom langchain_openai import ChatOpenAI
from pydantic import BaseModel, Fieldclass AnswerWithJustification(BaseModel):'''An answer to the user question along with justification for the answer.'''answer: strjustification: Optional[str] = Field(default=..., description="A justification for the answer.")from langchain_openai import ChatOpenAI
llm = ChatOpenAI(temperature=0,model="GLM-4-flash",openai_api_key="your api key",openai_api_base="https://open.bigmodel.cn/api/paas/v4/"
)
structured_llm = llm.with_structured_output(AnswerWithJustification, strict=True
)structured_llm.invoke("What weighs more a pound of bricks or a pound of feathers"
)

AnswerWithJustification(answer='a pound of bricks', justification="The weight of an object is measured in pounds, and a pound of bricks and a pound of feathers both weigh the same amount, one pound. The phrase 'a pound of bricks or a pound of feathers' is often used to illustrate the concept of weight versus volume, as bricks are denser and have less volume compared to feathers, making them heavier per unit volume.")

2. include_raw=True

from langchain_openai import ChatOpenAI
from pydantic import BaseModelclass AnswerWithJustification(BaseModel):'''An answer to the user question along with justification for the answer.'''answer: strjustification: strstructured_llm = llm.with_structured_output(AnswerWithJustification, include_raw=True
)structured_llm.invoke("What weighs more a pound of bricks or a pound of feathers"
)

{'raw': AIMessage(content='', additional_kwargs={'tool_calls': [{'id': 'call_-9158287991605179419', 'function': {'arguments': '{"answer": "a pound of bricks", "justification": "The weight of an object is measured in pounds, and a pound of bricks and a pound of feathers both weigh the same amount, one pound. The phrase \'a pound of bricks or a pound of feathers\' is often used to illustrate the concept of weight versus volume, as bricks are denser and take up less volume than feathers for the same weight."}', 'name': 'AnswerWithJustification'}, 'type': 'function', 'index': 0}], 'refusal': None}, response_metadata={'token_usage': {'completion_tokens': 92, 'prompt_tokens': 187, 'total_tokens': 279, 'completion_tokens_details': None, 'prompt_tokens_details': None}, 'model_name': 'GLM-4-flash', 'system_fingerprint': None, 'finish_reason': 'tool_calls', 'logprobs': None}, id='run-b26e455f-34b1-46bf-92b5-9a6be1aa4379-0', tool_calls=[{'name': 'AnswerWithJustification', 'args': {'answer': 'a pound of bricks', 'justification': "The weight of an object is measured in pounds, and a pound of bricks and a pound of feathers both weigh the same amount, one pound. The phrase 'a pound of bricks or a pound of feathers' is often used to illustrate the concept of weight versus volume, as bricks are denser and take up less volume than feathers for the same weight."}, 'id': 'call_-9158287991605179419', 'type': 'tool_call'}], usage_metadata={'input_tokens': 187, 'output_tokens': 92, 'total_tokens': 279, 'input_token_details': {}, 'output_token_details': {}}),'parsed': AnswerWithJustification(answer='a pound of bricks', justification="The weight of an object is measured in pounds, and a pound of bricks and a pound of feathers both weigh the same amount, one pound. The phrase 'a pound of bricks or a pound of feathers' is often used to illustrate the concept of weight versus volume, as bricks are denser and take up less volume than feathers for the same weight."),'parsing_error': None}

3. schema

from langchain_openai import ChatOpenAIoai_schema = {'name': 'AnswerWithJustification','description': 'An answer to the user question along with justification for the answer.','parameters': {'type': 'object','properties': {'answer': {'type': 'string'},'justification': {'description': 'A justification for the answer.', 'type': 'string'}},'required': ['answer']}
}structured_llm = llm.with_structured_output(oai_schema)structured_llm.invoke("What weighs more a pound of bricks or a pound of feathers"
)

{'answer': 'a pound of bricks','justification': "The weight of an object is measured in pounds, and a pound of bricks and a pound of feathers both weigh the same amount, one pound. The phrase 'a pound of bricks or a pound of feathers' is a riddle that plays on the fact that both weigh the same, but bricks are denser and heavier in volume compared to feathers. However, in terms of actual weight, they are equal."}

4. method=“json_mode”, include_raw=True

from langchain_openai import ChatOpenAI
from pydantic import BaseModelclass AnswerWithJustification(BaseModel):answer: strjustification: strstructured_llm = llm.with_structured_output(AnswerWithJustification,method="json_mode",include_raw=True
)structured_llm.invoke("Answer the following question. ""Make sure to return a JSON blob with keys 'answer' and 'justification'.\\n\\n""What's heavier a pound of bricks or a pound of feathers?"
)

{'raw': AIMessage(content='{"answer":"a pound of bricks","justification":"The question is a classic riddle. The answer is \\u0027a pound of bricks\\u0027 because both the bricks and the feathers weigh one pound, but bricks are denser and thus heavier in terms of mass. The riddle plays on the word \\u0027heavier,\\u0027 which can refer to mass rather than weight."}', additional_kwargs={'parsed': None, 'refusal': None}, response_metadata={'token_usage': {'completion_tokens': 85, 'prompt_tokens': 97, 'total_tokens': 182, 'completion_tokens_details': None, 'prompt_tokens_details': None}, 'model_name': 'GLM-4-flash', 'system_fingerprint': None, 'finish_reason': 'stop', 'logprobs': None}, id='run-5532c43c-45ad-4ae3-8a4f-ea2eea980778-0', usage_metadata={'input_tokens': 97, 'output_tokens': 85, 'total_tokens': 182, 'input_token_details': {}, 'output_token_details': {}}),'parsed': AnswerWithJustification(answer='a pound of bricks', justification="The question is a classic riddle. The answer is 'a pound of bricks' because both the bricks and the feathers weigh one pound, but bricks are denser and thus heavier in terms of mass. The riddle plays on the word 'heavier,' which can refer to mass rather than weight."),'parsing_error': None}

查看全文

http://www.mrgr.cn/news/80160.html