Lewati ke konten utama

Conversation Flow

Overview​

Conversation Flow is the dynamic conversational pathway between a user and an agent in the Perfect10 platform. This system is designed to provide contextual, personalized, and relevant responses based on user input and ongoing conversation memory.

Conversation Architecture​

Main Components​

1. Intent & Entity Recognition​

  • The system identifies the intent behind the user's input.
  • Detects key entities such as product names, dates, numbers, locations, etc.

2. Memory (Short-Term & Long-Term)​

  • Short-Term Memory: Context from the most recent messages.
  • Long-Term Memory: Explicitly stored key facts, such as user preferences, ticket IDs, etc.

3. RAG (Retrieval-Augmented Generation)​

  • The system retrieves additional information from the Knowledge Base when needed.
  • Can be triggered automatically or by the agent.

4. Response Generation​

  • Uses an LLM (e.g., Gemini, GPT-4) to compose responses based on input, memory, and relevant documents.

5. Conversation Routing (Optional)​

  • Redirects the conversation to a different agent or a human if needed.
  • Used in complex workflows (e.g., customer service escalation, internal approvals, etc.).

Conversation Modes​

Perfect10 supports several conversation types:

ModeDescription
Open ChatFreeform conversation without structural constraints
Form-likeStructured interaction, e.g., filling out forms, surveys, or registrations
Guided FlowUses buttons, quick replies, and predefined scenarios
Multi-turn ReasoningEnables the system to address complex questions step-by-step

Example Conversation Flow​

Case 1: Customer wants to check delivery status​

  1. User: "I want to know the status of my order."
  2. Agent: "Could you provide your order number?"
  3. User: "INV-3921"
  4. Agent:
    • Checks the system via API
    • Responds: "Your order INV-3921 is currently being packaged and will be shipped tomorrow."

Case 2: Financial calculation using Calculator Tools​

  1. User: "I need to calculate the compound interest for my investment."
  2. Agent: "I can help you with that! Please provide the principal amount, interest rate, and time period."
  3. User: "Principal: $10,000, Rate: 8% per year, Time: 5 years"
  4. Agent:
    • Uses Calculator Tools: 10000 * (1 + 0.08)^5
    • Responds: "Your investment will grow to $14,693.28 after 5 years with compound interest."

Case 3: Statistical analysis using Calculator Tools​

  1. User: "What's the average score of my students?"
  2. Agent: "Please provide the student scores."
  3. User: "Scores: 85, 92, 78, 88, 95, 82"
  4. Agent:
    • Uses Calculator Tools: (85 + 92 + 78 + 88 + 95 + 82) / 6
    • Responds: "The average score is 86.67."

Action Integration (Agent Action)​

Agents can perform specific actions, such as:

  • Mathematical Calculations: Using Calculator Tools for financial analysis, statistical computations
  • Web Search: Querying external APIs for real-time information
  • URL Access: Retrieving content from web pages
  • Data Processing: Saving data to a database, triggering internal workflows
  • API Integration: Executing plugins or commands through external services

Tips for Designing Conversation Flows​

  • Use a consistent agent persona.
  • Create fallback responses for unrecognized inputs.
  • Avoid asking too many questions in a single response.
  • Implement input validation in structured chat modes.
  • Consider using quick replies for faster interactions.

Logging & Analysis​

Every conversation flow is logged in detail:

  • User and agent messages
  • Documents used (if RAG)
  • Actions executed
  • Response time and token usage
  • Session ID for tracking