Dialogue Agent

Author: svcvit Version: 0.0.3

Description

A powerful task-oriented dialogue agent that can collect information through structured conversations. It supports dynamic field validation, multi-field information extraction, and state management. The agent now includes two dialogue strategies: TOD (Task-Oriented Dialogue) for information collection and MTD (Multi-Turn Dialogue) for general conversations.

Features

• Task-oriented dialogue management
• Dynamic field validation
• Multi-field information extraction
• Conversation state persistence
• Automatic answer validation
• Context-aware information collection
• Natural language interaction
• Multiple dialogue strategies support (TOD & MTD)

Usage Guide

Dialogue Strategies

The agent supports two dialogue strategies:

1. TOD (Task-Oriented Dialogue)

   • A dialogue strategy for structured information collection
   • Suitable for form filling and information gathering scenarios
   • Supports dynamic field validation and state management

2. MTD (Multi-Turn Dialogue)

   • A strategy for general multi-turn conversations
   • Supports context awareness
   • Suitable for open-ended dialogue scenarios

TOD Strategy Parameters

1. information_schema (Required)

   • Type: string (JSON)
   • Description: Schema defining the fields to collect
   • Example:

     {
       "fields": [
         {
           "name": "destination",
           "question": "请问您想去哪里旅行？",
           "required": true
         },
         {
           "name": "duration",
           "question": "您计划旅行多长时间？",
           "required": true
         },
         {
           "name": "budget",
           "question": "您的预算大约是多少？",
           "required": true
         }
       ]
     }

2. query (Required)

   • Type: string
   • Description: User's input text

3. model (Required)

   • Type: AgentModelConfig
   • Description: LLM model configuration
   • Example: Configuration for GPT or other LLM models

4. storage_key (Required)

   • Type: string
   • Description: Unique key for storing conversation state
   • Example: "conversation-123"

MTD Strategy Parameters

1. instruction (Required)

   • Type: string
   • Description: Dialogue instruction or system prompt
   • Example:

      你是一个专业的医生。
      - 问用户有什么不适症状
      - 问用户最近有没有吃药
      - 问如果用户回答吃药了，询问吃的什么药，以及剂量。
      - 最后询问用户有没有药物过敏。
      - 最后询问用户有没有其他症状。
     

2. query (Required)

   • Type: string
   • Description: User's input text

3. model (Required)

   • Type: AgentModelConfig
   • Description: LLM model configuration
   • Example: Configuration for GPT or other LLM models

4. storage_key (Required)

   • Type: string
   • Description: Unique key for storing conversation state
   • Example: "mtd-conversation-123"

Changelog

v0.0.3

• Added MTD (Multi-Turn Dialogue) strategy support
• Fixed data type validation bugs
• Added support for schema JSON without leading spaces (compatible with Dify official update)

v0.0.2

• Added logging functionality
• Added token usage statistics
• Optimized code structure and performance

v0.0.1

• Project initialization
• Implemented multi-turn dialogue
• Implemented conversation state storage
• Implemented intelligent Q&A content extraction