JSON 代理

本文档展示了一个专为与大型 JSON/dict 对象交互而设计的代理。当您想要回答关于一个太大以至于无法适应 LLM 上下文窗口的 JSON 数据块的问题时，这将非常有用。代理能够迭代地探索数据块，以找到答案所需的信息。

在下面的示例中，我们使用了 OpenAI API 的 OpenAPI 规范，您可以在这里找到它。

我们将使用 JSON 代理来回答关于 API 规范的一些问题。

初始化

import os
import yaml

from langchain.agents import (
    create_json_agent,
    AgentExecutor
)
from langchain.agents.agent_toolkits import JsonToolkit
from langchain.chains import LLMChain
from langchain.llms.openai import OpenAI
from langchain.requests import TextRequestsWrapper
from langchain.tools.json.tool import JsonSpec

with open("openai_openapi.yml") as f:
    data = yaml.load(f, Loader=yaml.FullLoader)
json_spec = JsonSpec(dict_=data, max_value_length=4000)
json_toolkit = JsonToolkit(spec=json_spec)

json_agent_executor = create_json_agent(
    llm=OpenAI(temperature=0),
    toolkit=json_toolkit,
    verbose=True
)

示例：获取请求的必需 POST 参数

json_agent_executor.run("What are the required parameters in the request body to the /completions endpoint?")

Entering new AgentExecutor chain...
Action: json_spec_list_keys
Action Input: data
Observation: ['openapi', 'info', 'servers', 'tags', 'paths', 'components', 'x-oaiMeta']
Thought: I should look at the paths key to see what endpoints exist
Action: json_spec_list_keys
Action Input: data["paths"]
Observation: ['/engines', '/engines/{engine_id}', '/completions', '/edits', '/images/generations', '/images/edits', '/images/variations', '/embeddings', '/engines/{engine_id}/search', '/files', '/files/{file_id}', '/files/{file_id}/content', '/answers', '/classifications', '/fine-tunes', '/fine-tunes/{fine_tune_id}', '/fine-tunes/{fine_tune_id}/cancel', '/fine-tunes/{fine_tune_id}/events', '/models', '/models/{model}', '/moderations']
Thought: I should look at the /completions endpoint to see what parameters are required
Action: json_spec_list_keys
Action Input: data["paths"]["/completions"]
Observation: ['post']
Thought: I should look at the post key to see what parameters are required
Action: json_spec_list_keys
Action Input: data["paths"]["/completions"]["post"]
Observation: ['operationId', 'tags', 'summary', 'requestBody', 'responses', 'x-oaiMeta']
Thought: I should look at the requestBody key to see what parameters are required
Action: json_spec_list_keys
Action Input: data["paths"]["/completions"]["post"]["requestBody"]
Observation: ['required', 'content']
Thought: I should look at the required key to see what parameters are required
Action: json_spec_get_value
Action Input: data["paths"]["/completions"]["post"]["requestBody"]["required"]
Observation: True
Thought: I should look at the content key to see what parameters are required
Action: json_spec_list_keys
Action Input: data["paths"]["/completions"]["post"]["requestBody"]["content"]
Observation: ['application/json']
Thought: I should look at the application/json key to see what parameters are required
Action: json_spec_list_keys
Action Input: data["paths"]["/completions"]["post"]["requestBody"]["content"]["application/json"]
Observation: ['schema']
Thought: I should look at the schema key to see what parameters are required
Action: json_spec_list_keys
Action Input: data["paths"]["/completions"]["post"]["requestBody"]["content"]["application/json"]["schema"]
Observation: ['$ref']
Thought: I should look at the $ref key to see what parameters are required
Action: json_spec_get_value
Action Input: data["paths"]["/completions"]["post"]["requestBody"]["content"]["application/json"]["schema"]["$ref"]
Observation: #/components/schemas/CreateCompletionRequest
Thought: I should look at the CreateCompletionRequest schema to see what parameters are required
Action: json_spec_list_keys
Action Input: data["components"]["schemas"]["CreateCompletionRequest"]
Observation: ['type', 'properties', 'required']
Thought: I should look at the required key to see what parameters are required
Action: json_spec_get_value
Action Input: data["components"]["schemas"]["CreateCompletionRequest"]["required"]
Observation: ['model']
Thought: I now know the final answer
Final Answer: The required parameters in the request body to the /completions endpoint are 'model'.

Finished chain.

"The required parameters in the request body to the /completions endpoint are 'model'."