Agent 设计模式——multiple tools

const META_AGENT_PROMPT = `
You are APIGPT, with access to a database of APIs. This database is organized
into the following categories: {categories}. Your task is to help users
identify the relevant categories for their needs. To do this, you can
use the 'get_tools_in_category' function to retrieve the available tools
within a specific category. If you are unsure about the functionality of
some tools, the 'get_tools_descriptions' function can be used to obtain
detailed information about these tools. This information will aid you in
understanding the general functionality of each category. Additionally, the
'create_agent_category_level' function allows you to assign a relevant category
to an agent, with each agent being assigned only one category. However,
you can assign multiple categories to different agents. It is important
to explore as many categories as possible, as the solution to a query may
be found in unexpected categories. Remember, your goal is not to answer
the query directly but to identify all potentially relevant categories and
assign them to agents. Once you have completed the assignment, call the
'Finish' function.
At each step,  you should call functions to actually excute your step.
All the thought is short, at most in 3 sentence.
`;
 
const CATEGORY_AGENT_PROMPT = `
You are APIGPT, with access to a database of APIs categorized into various
groups. Each category contains numerous tools, and each tool encompasses
multiple APIs. Your task is to assist users in finding relevant tools within
the category: {category}. If uncertain about the functionality of some tools, use
the 'get_tools_descriptions' function to obtain detailed information. Then,
employ the 'create agent tool level' function to allocate a subset of pertinent
tools to an agent, ensuring that similar tools are assigned to the same agent
and limiting the allocation to no more than five tools per agent. You may
assign different subsets to multiple agents. Remember, your role is not to
answer queries directly, but to assign all possible tools. Once you complete
the assignment, or if you determine the query is irrelevant to the tools in
the specified category, invoke the 'Finish' function.
At each step,  you should call functions to actually excute your step.
All the thought is short, at most in 3 sentence.
`;
 
const TOOL_AGENT_PROMPT = `
You are APIGPT with access to a database of APIs, categorized into various
sections. Each category contains multiple tools, and each tool encompasses
numerous APIs. Your task is to assist users in finding relevant APIs within
the tools '{tools}' of the '{category}' category. You will be provided with
descriptions and details of these tools and their APIs. Upon identifying
relevant API names, use the 'add_apis_into_api_pool' function to add them to
the final API list. If you conclude that all possible APIs have been explored,
or if there are no relevant APIs in these tools, invoke the Finish function.
During the process, you may receive feedback on these APIs.
At each step,  you should call functions to actually excute your step.
All the thought is short, at most in 3 sentence.
`;

const REFIND_CATEGORY_PROMPT = `
Current APIs failed to solve the query and the result is: {{failed_reason}}.
Please assign more unexplored categories to the agents.
`;
 
const REFIND_TOOL_PROMPT = `
Current APIs failed  to solve the query. The result is: {{failed_reason}}.
Please assign more unexplored tools to the agents.
`;
 
const REFIND_API_PROMPT = `
Current APIs failed to solve the query. The result is: {{failed_reason}}.
You need to analyze the result, and find more apis.
It is possible that the tools do not have the relevant apis. In this case, you should call the Finish function. Do not make up the tool names or api names.
`;

export const CATEGORY_MAPPING = {
  Data_and_Analytics: [
    "Data",
    "Database",
    "Text_Analysis",
  ],
  Travel_and_Transportation: ['Travel', 'Transportation', 'Logistics', 'Location'],
  //...
};
 
export const apiList = [{
   {
    "id": "c84af6be-804a-4965-93a5-cdbc29586f00",
    "category_name": "Travel",
    "tool_name": "Booking com",
    "api_name": "Hotels Search",
    "api_description": "string",
    "required_parameters": [
      {
        "name": "string",
        "type": "STRING",
        "description": "",
        "default": "popularity"
      },
    ],
    "optional_parameters": [],
    "method": "GET",
    "api_url": "https://booking-com.p.rapidapi.com/v2/hotels/search"
  },
}]
 
// 用户问题
const query = `I'm organizing a charity event to raise awareness for animal rights. Can you recommend a book that highlights the importance of compassion towards animals? Additionally, provide me with a random word that symbolizes unity and empathy`

const graphChannels = {
  // llm 实例
  llm: null,
  // 用户查询内容
  query: null,
  // 匹配类别
  categories: null,
  // 该类别下对应 api 列表
  apis: null,
  // 模型返回最佳匹配 api
  bestApi: null,
  // 模型从 query 提取参数
  params: null,
  // 执行 api 返回结果
  response: null,
};

// 1. LLM 从用户提问中提取分类
graph.addNode("extract_category_node", extractCategory);
 
// 2. 从前一节点提取到 category 下的所有 api 列表
graph.addNode("get_apis_node", getApis);
 
// 3. 选择最合适的 api
graph.addNode("select_api_node", selectApi);
 
// 4. 获取 api 所需参数
graph.addNode("extract_params_node", extractParameters);
 
// 5. 补充 api 缺失必需参数
graph.addNode("human_loop_node", requestParameters);
 
// 6. 执行请求，获取最终结果
graph.addNode("execute_request_node", createFetchRequest);

// 1. 起点 => 提取分类节点
graph.addEdge(START, "extract_category_node");
 
// 2. 分类节点 => 分类接口接口
graph.addEdge("extract_category_node", "get_apis_node");
 
// 3. 分类节点 => 匹配api 节点
graph.addEdge("get_apis_node", "select_api_node");
 
// 4. 匹配 api 节点 => 抽取 api 参数节点
graph.addEdge("select_api_node", "extract_params_node");
 
// 条件判断连线
graph.addConditionalEdges("extract_params_node", verifyParams);
graph.addConditionalEdges("human_loop_node", verifyParams);
 
// 5. 完成 API 请求，循环结束
graph.addEdge("execute_request_node", END);
 
const verifyParams = (
  state: GraphState
): "human_loop_node" | "execute_request_node" => {
  const { bestApi, params } = state;
  // 参数校验
  const missingKeys = findMissingParams(
    requiredParamsKeys,
    extractedParamsKeys
  );
  // 如果必需参数校验不通过，先补充参数
  if (missingKeys.length > 0) {
    return "human_loop_node";
  }
  // 校验通过，执行请求
  return "execute_request_node";
};