Skip to content

Latest commit

 

History

History
102 lines (82 loc) · 3.93 KB

README.md

File metadata and controls

102 lines (82 loc) · 3.93 KB

LiteAgent core for dart

English · 中文

LLM AI Agent multi sessions service.

Features

  • Support pure text agent without JSON Spec.
  • Support OpenAPI/OpenRPC/OpenModbus/OpenTool JSON Spec.
  • Support LLM Function calling to HTTP API/json-rpc 2.0 over HTTP/Modbus and more custom tools.

Usage

Prepare

  1. Some OpenSpec json file, according to /example/json/open*/*.json, which is callable.
  2. Run your tool server, which is described in json file.
  3. Add .env file in the example folder, and add below content in the .env file： 
    baseUrl = https://xxx.xxx.com         # LLM API BaseURL
apiKey = sk-xxxxxxxxxxxxxxxxxxxx      # LLM API ApiKey
  4. Use below method to run agent service.

Method 1(Recommend): AgentService

  • According to /example/service_example
  • Support multi agent session via session id.
  • Support multi task in the same agent, identify different tasks by taskId. After finishing task, task message could be added to session as new task context.
Future<void> main() async {
  CapabilityDto capabilityDto = CapabilityDto(
      llmConfig: _buildLLMConfig(),       // LLM Config
      systemPrompt: _buildSystemPrompt(), // System Prompt
      openSpecList: await _buildOpenSpecList()  // OpenSpec Description String List
  );
  SessionDto sessionDto = await agentService.initChat(
      capabilityDto,
      listen        // Subscribe AgentMessage, Agent chat with User/Client/LLM/Tools Role
  ); // Get Session Id
  
  String prompt = "<USER PROMPT, e.g. call any one tool>";
  UserTaskDto userTaskDto = UserTaskDto(taskId: "<Identify different tasks, NOT more than 36 chars>", contentList: [UserMessageDto(type: UserMessageDtoType.text, message: prompt)]);  // User Content List, support type text/imageUrl
  await agentService.startChat(sessionDto.id, userTaskDto);
}
  • MultiAgent support
Future<void> main() async {
  SessionDto sessionDto1 = await _buildTextAgent();
  SessionDto sessionDto2 = await _buildToolAgent();

  CapabilityDto capabilityDto = CapabilityDto(llmConfig: llmConfig, systemPrompt: systemPrompt,
      sessionList: [sessionDto1, sessionDto2]
  );

  SessionDto sessionDto = await agentService.initChat(capabilityDto, listen);

  UserTaskDto userTaskDto = UserTaskDto(contentList: [UserMessageDto(type: UserMessageDtoType.text, message: prompt)]);
  await agentService.startChat(sessionDto.id, userTaskDto);
}
  • Reflection support
Future<void> main() async {
  CapabilityDto capabilityDto = CapabilityDto(
      llmConfig: _buildLLMConfig(),
      systemPrompt: _buildSystemPrompt(),
      openSpecList: await _buildOpenSpecList(),
      /// Add reflection prompt list here
      toolReflectionList: await _buildToolReflectionList()
  );
  SessionDto sessionDto = await agentService.initChat(capabilityDto, listen);
  String prompt = "<USER PROMPT, e.g. call any one tool>";
  UserTaskDto userTaskDto = UserTaskDto(taskId: "<Identify different tasks, NOT more than 36 chars>", contentList: [UserMessageDto(type: UserMessageDtoType.text, message: prompt)]);
  await agentService.startChat(sessionDto.id, userTaskDto);
}

Method 2: ToolAgent

  • According to /example/agent_example
  • Pure native calling. Support single session.
  • Method 1 AgentService is friendly encapsulation for this.
Future<void> main() async {
  ToolAgent toolAgent = ToolAgent(
      llmRunner: _buildLLMRunner(),
      session: _buildSession(),
      toolRunnerList: await _buildToolRunnerList(),
      systemPrompt: _buildSystemPrompt()
  );
  String prompt = "<USER PROMPT, e.g. call any one tool>";
  toolAgent.userToAgent(taskId: "<Identify different tasks, NOT more than 36 chars>", [Content(type: ContentType.text, message: prompt)]);
}