customer_support.json

Overview

The customer_support.json file defines a comprehensive intelligent customer service workflow system. Its primary function is to process user queries by classifying their intent via a large language model (LLM) and directing the requests to specialized agents for tailored responses. This system supports multilingual titles and descriptions (English and Chinese) and is designed to simulate a professional, empathetic, and context-aware customer support experience.

The workflow is structured as a directed graph of interconnected components including:

• Begin component: Entry point that initiates the conversation.

• Categorize component: Classifies user queries into predefined categories (contact, casual, complain, product-related).

• Retrieval component: Fetches relevant product information from a knowledge base.

• Agent components: Specialized chatbots (agents) that handle different query categories with unique behaviors and LLM prompts.

• Message components: Generate final responses by aggregating agent outputs or providing fixed replies.

This modular design enables flexible, multi-path processing of user intents and ensures that queries are handled by the most appropriate expert agent.

Detailed Component Descriptions

1. Begin Component

• Name: begin

• Type: Begin

• Role: Entry point of the workflow.

• Parameters:

  • enablePrologue (bool): Enables an introductory message.

  • prologue (string): Initial message to greet users.

  • mode (string): Conversation mode, set to "conversational".

• Functionality: Starts the interaction with a welcoming prompt:
  "Hi! I'm an official AI customer service representative. How can I help you?"

• Downstream: Forwards user input to the Categorize component for intent classification.

2. Categorize Component

• Name: Categorize:DullFriendsThank

• Type: Categorize

• Role: Classifies user queries into four categories using LLM intent classification.

• Parameters:

  • category_description: Defines four categories with descriptions, examples, and routing targets:

    • contact: Queries requesting specific contact info (email, phone, social handles).

    • casual: Non-product related casual chat.

    • complain: General complaints or dissatisfaction.

    • product related: Questions about product usage, features, issues.

  • llm_id: LLM model identifier (deepseek-chat@DeepSeek).

  • temperature: Controls LLM response randomness (0.1).

  • message_history_window_size: 1 (only recent message considered).

  • outputs: category_name of type string.

  • query: Bound to system query sys.query.

• Functionality: Uses the LLM to assign the user query to one of the four categories.

• Downstreams: Routes queries to respective components based on category:

  • contact → Message:BreezyDonutsHeal

  • casual → Agent:TwelveOwlsWatch

  • complain → Agent:DullTownsHope

  • product related → Retrieval:ShyPumasJoke

3. Retrieval Component

• Name: Retrieval:ShyPumasJoke

• Type: Retrieval

• Role: Retrieves relevant product document snippets from knowledge bases in response to product-related queries.

• Parameters:

  • query: System query (sys.query).

  • top_k: 1024 (max documents retrieved).

  • top_n: 8 (top documents used).

  • similarity_threshold: 0.2 (minimum similarity to consider).

  • keywords_similarity_weight: 0.7 (weight for keyword similarity in retrieval).

  • empty_response: Empty string if no docs found.

  • outputs: formalized_content - a string containing retrieved content.

• Functionality: Fetches product-related information that is then used by a dedicated product agent for answering.

• Downstream: Passes retrieved content to Agent:KhakiSunsJudge.

4. Agent Components

Each Agent represents a specialized chatbot with distinct roles, system prompts, and LLM settings.

a) Agent: DullTownsHope ("Soothe mood")

• Role: Provides empathetic, emotional support to users who complain or express frustration.

• Prompts / Behavior:

  • Warm, kind, understanding tone.

  • Focus on empathy rather than solving the problem.

  • Encourages users positively.

• LLM Parameters:

  • Model: deepseek-chat@DeepSeek.

  • Temperature: 0.5.

  • Max tokens: 4096.

  • Max retries: 3.

  • Max rounds: 5.

  • Frequency penalty: 0.3.

  • Presence penalty: 0.2.

• Inputs: User query (sys.query).

• Downstream: Sends output to Message:GreatDucksArgue.

b) Agent: KhakiSunsJudge ("Product info")

• Role: Acts as a highly professional product information advisor, providing strictly factual and structured answers based on retrieved documents.

• Prompts / Behavior:

  • Answers only from verified product documents.

  • No assumptions, guesses, or fabricated content.

  • Requests clarification if query is unclear.

  • Answers are concise, professional, and formatted.

• LLM Parameters:

  • Model: deepseek-chat@DeepSeek.

  • Temperature: 0.1.

  • Max tokens: 256.

  • Frequency penalty: 0.7.

  • Presence penalty: 0.4.

  • Max retries: 3.

  • Max rounds: 5.

• Inputs: User query + retrieved product content (Retrieval:ShyPumasJoke@formalized_content).

• Downstream: Sends output to Message:GreatDucksArgue.

c) Agent: TwelveOwlsWatch ("Causal chat")

• Role: Engages the user in friendly, casual conversation.

• Prompts / Behavior:

  • Natural, relaxed, positive tone.

  • Avoids sensitive or negative topics.

  • May guide conversation with related casual topics.

• LLM Parameters:

  • Model: deepseek-chat@DeepSeek.

  • Temperature: 0.5.

  • Max tokens: 4096.

  • Frequency penalty: 0.3.

  • Presence penalty: 0.2.

• Inputs: User query (sys.query).

• Downstream: Sends output to Message:GreatDucksArgue.

5. Message Components

a) Message: BreezyDonutsHeal

• Role: Sends fixed responses for contact category queries.

• Parameters: A list of canned replies such as:

  • "Okay, I've already write this down. What else I can do for you?"

  • "Thanks for your trust! Our expert will contact ASAP. So, anything else I can do for you?"

• Downstream: None (endpoint).

• Upstream: Receives from Categorize:DullFriendsThank.

b) Message: GreatDucksArgue ("Response")

• Role: Aggregates all agent responses for complain, casual, and product-related queries into a combined final message.

• Parameters: Concatenates outputs from all three agents:

  • {Agent:TwelveOwlsWatch@content}{Agent:DullTownsHope@content}{Agent:KhakiSunsJudge@content}

• Downstream: None (endpoint).

• Upstream: Receives from all three agent components.

Global Variables

• sys.conversation_turns: Tracks number of conversation turns.

• sys.files: List of files involved (empty by default).

• sys.query: The current user query string.

• sys.user_id: ID of the current user.

Workflow / Execution Flow

1. Conversation Start: The begin node greets the user and collects their query.

2. Intent Classification: The Categorize node classifies the user's intent into one of four categories.

3. Routing:

   • If contact, respond with fixed messages (Message:BreezyDonutsHeal).

   • If casual, forward query to casual chat agent (Agent:TwelveOwlsWatch).

   • If complain, forward to empathetic agent (Agent:DullTownsHope).

   • If product related, retrieve relevant documents (Retrieval:ShyPumasJoke), then forward to product info agent (Agent:KhakiSunsJudge).

4. Response Generation: Agents generate responses based on their respective roles and LLM prompts.

5. Response Aggregation: Message:GreatDucksArgue aggregates outputs from the agents for final user output.

6. End of Turn: The system awaits the next user input or ends the session.

Implementation Details and Algorithms

• Intent Classification: Uses an LLM (deepseek-chat@DeepSeek) with prompt engineering to classify incoming queries into well-defined categories. The categories are clearly described with example queries for better classification accuracy.

• Product Information Retrieval: Implements a vector-based semantic retrieval system that fetches relevant product documents using similarity scoring with keyword weighting and thresholds. This preprocessed content is passed to the product agent for fact-based answering.

• Agent Responses: Each agent has a finely tuned system prompt that shapes the LLM output to match the intended tone and role:

  • Empathetic assistance for complaints.

  • Professional, fact-based answers for product queries.

  • Lighthearted, casual conversation for chit-chat.

• Response Aggregation: Multiple agents' outputs are concatenated to provide a comprehensive final user response, accommodating multi-intent or overlapping queries.

• Error Handling: Agents have retry limits and delay parameters for robust error recovery.

Interaction with Other System Components

• Knowledge Base: The Retrieval component interacts with external product knowledge bases to fetch relevant documents.

• LLM Services: All classification and response generation rely on the specified LLM deepseek-chat@DeepSeek.

• User Interface Layer: The system expects a frontend or conversational interface to deliver prologue, collect sys.query, and display aggregated responses.

• System Globals: Shared variables like sys.query and sys.conversation_turns maintain context across components.

Usage Example

Suppose a user inputs the query:

"How do I install the new model X100 on the wall?"

1. begin greets and receives the query.

2. Categorize classifies it as product related.

3. Retrieval fetches relevant product documents related to "model X100 installation."

4. Agent:KhakiSunsJudge uses the retrieved info and the query to compose an accurate, factual answer.

5. The answer is combined with outputs from other agents (empty in this case) in Message:GreatDucksArgue.

6. The final response is shown to the user.

Visual Diagram: Component Interaction Flowchart

flowchart LR
    Begin["Begin\n(Start Conversation)"]
    Categorize["Categorize\n(Intent Classification)"]
    ContactMsg["Message: BreezyDonutsHeal\n(Contact Replies)"]
    CasualAgent["Agent: TwelveOwlsWatch\n(Casual Chat)"]
    ComplainsAgent["Agent: DullTownsHope\n(Empathetic Support)"]
    Retrieval["Retrieval: ShyPumasJoke\n(Product Info Retrieval)"]
    ProductAgent["Agent: KhakiSunsJudge\n(Product Info Advisor)"]
    FinalMsg["Message: GreatDucksArgue\n(Response Aggregation)"]

    Begin --> Categorize

    Categorize -- contact --> ContactMsg

    Categorize -- casual --> CasualAgent
    Categorize -- complain --> ComplainsAgent
    Categorize -- product related --> Retrieval

    Retrieval --> ProductAgent

    CasualAgent --> FinalMsg
    ComplainsAgent --> FinalMsg
    ProductAgent --> FinalMsg

Summary

The customer_support.json file describes a modular, LLM-powered intelligent customer support workflow that classifies user intents and routes queries to specialized agents for empathetic, casual, contact-related, or product information responses. The workflow integrates a knowledge retrieval system and multiple finely tuned LLM prompts to ensure accurate, context-aware, and user-friendly interactions.

This file serves as the core orchestration layer for an AI-driven customer service system, enabling scalable and maintainable conversational experiences.

End of Documentation for customer_support.json