Multi-Turn Conversation

Overview

Flow ID: multi-turn-conversation
Category: Chat Interactions
Estimated Duration: Variable
User Role: All Users
Complexity: Simple

Purpose: Describes the AI’s ability to maintain “memory” of previous exchanges within a chat session. This allows users to ask follow-up questions (“What about the second point?”) without repeating context.

Trigger

What initiates this flow:

• User follows up on a previous answer

Prerequisites

Before starting, users must have:

• An active chat with at least one message exchange

User Intent Analysis

Primary Intent

Refine, expand, or correct the AI’s previous output through natural dialogue.

Step-by-Step Flow

Main Path (Happy Path)

Step 1: Initial Question

• User Action: “List 3 fruits.”
• AI Response: “1. Apple, 2. Banana, 3. Cherry.”

Step 2: Follow-Up (Implicit Context)

• User Action: “Which one is red?” (User doesn’t say “fruit”).
• System Action: Sends [History + New Prompt] to LLM.
• AI Response: “Both Apples and Cherries are typically red.” (Correctly resolving reference).

Step 3: Context Limit Handling

• Condition: Conversation exceeds Context Window (e.g., > 4096 tokens).
• System Action: “Sliding Window” - Oldest messages are dropped from the prompt sent to LLM.
• User Experience: AI might “forget” the very first thing said if chat is very long.

Error States & Recovery

Error 1: AI Forgets Context

Cause: Chat history exceeded token limit
User Experience: User asks “What was the date again?” and AI says “I don’t know what date you mean.”
Recovery: User must restate key facts. Or increase Context Window Size.

Design Considerations

• History Pruning: Intelligently summarize old history rather than just hard-deleting it (if supported).
• Visuals: Auto-scroll to bottom on new message.

Related Flows

• New Chat Empty
• Configure Context Window

Technical References

• src/engines/chat-memory.js
• src/constants/model-params.js