Multi-Turn Conversations

What are Multi-Turn Conversations?

Multi-turn conversations are interactions where the AI remembers previous messages and maintains context:

User: "I need a laptop"
AI: "What's your budget?"

User: "$1000"                          ← References "laptop" and "budget" from before
AI: "Here are laptops under $1000"     ← Remembers user needs laptop for ~$1000

The AI must:

✅ Remember previous messages
✅ Understand references (“it”, “that”, “the one you mentioned”)
✅ Maintain conversation state (user’s budget, preferences, etc.)
✅ Stay contextually relevant

Why Test Multi-Turn Conversations?

Single-turn testing misses critical issues:

Context Loss

Turn 1: "Find flights to Paris"
Turn 2: "What about hotels?"

❌ Without context: AI asks "Where?"
✅ With context: AI knows user means Paris

Reference Resolution

Turn 1: "Show me the MacBook Pro and Dell XPS"
Turn 2: "What's the battery life on the second one?"

❌ Without context: AI confused
✅ With context: AI knows "second one" = Dell XPS

State Management

Turn 1: "I want to book a flight"
Turn 2: "From Boston to Seattle"
Turn 3: "Next Friday"

AI must accumulate: flight booking, BOS→SEA, date=next Friday

Testing Approaches

Approach 1: Stateless (Recommended)

Test each conversation turn independently with full history:

{
  "tests": [
    {
      "id": "turn-1",
      "pipeline": [
        {
          "block": "HttpRequest",
          "input": {
            "body": {
              "messages": [
                { "role": "user", "content": "I need a laptop" }
              ]
            }
          }
        }
      ]
    },
    {
      "id": "turn-2",
      "pipeline": [
        {
          "block": "HttpRequest",
          "input": {
            "body": {
              "messages": [
                { "role": "user", "content": "I need a laptop" },
                { "role": "assistant", "content": "What's your budget?" },
                { "role": "user", "content": "$1000" }
              ]
            }
          }
        }
      ]
    }
  ]
}

Pros:

✅ Each test is independent
✅ Can run tests in parallel
✅ Easy to debug individual turns
✅ No shared state between tests

Cons:

❌ More verbose (repeat full history)
❌ Doesn’t test actual conversation flow

Approach 2: Stateful

Test conversation as continuous flow:

{
  "tests": [{
    "id": "full-conversation",
    "pipeline": [
      {
        "id": "turn-1",
        "block": "HttpRequest",
        "input": {
          "body": {
            "messages": [{ "role": "user", "content": "I need a laptop" }]
          }
        },
        "output": "response1"
      },
      {
        "block": "JsonParser",
        "input": "${response1.body}",
        "output": { "parsed": "data1" }
      },
      {
        "id": "turn-2",
        "block": "HttpRequest",
        "input": {
          "body": {
            "messages": [
              { "role": "user", "content": "I need a laptop" },
              { "role": "assistant", "content": "${data1.message}" },
              { "role": "user", "content": "$1000" }
            ]
          }
        },
        "output": "response2"
      }
    ]
  }]
}

Pros:

✅ Tests actual conversation flow
✅ Validates state management
✅ Closer to real usage

Cons:

❌ Tests are sequential (slower)
❌ Hard to isolate failures
❌ One failure breaks entire test

Using History in LLMJudge

LLMJudge accepts conversation history to evaluate responses in context:

{
  "block": "LLMJudge",
  "input": {
    "text": "${aiResponse}",
    "history": [
      { "role": "user", "content": "I'm looking for a laptop" },
      { "role": "assistant", "content": "What's your budget?" },
      { "role": "user", "content": "Around $1000" }
    ],
    "expected": {
      "expectedBehavior": "Recommends laptops in the $800-1200 range, remembering user's $1000 budget"
    }
  },
  "output": "judgement"
}

Without history, LLMJudge wouldn’t know the “$1000 budget” context.

Real-World Examples

1. Customer Service Bot

{
  "name": "Customer Service Conversation",
  "context": {
    "CHAT_URL": "${env.CHAT_API_URL}"
  },
  "tests": [
    {
      "id": "turn-1-greeting",
      "pipeline": [
        {
          "block": "HttpRequest",
          "input": {
            "url": "${CHAT_URL}/chat",
            "method": "POST",
            "body": {
              "messages": [
                { "role": "user", "content": "Hello" }
              ]
            }
          },
          "output": "response"
        },
        {
          "block": "JsonParser",
          "input": "${response.body}",
          "output": { "parsed": "data" }
        },
        {
          "block": "LLMJudge",
          "input": {
            "text": "${data.message}",
            "expected": {
              "expectedBehavior": "Friendly greeting and asks how they can help"
            }
          },
          "output": "judgement"
        }
      ],
      "assertions": {
        "judgement.score": { "gte": 0.8 }
      }
    },
    {
      "id": "turn-2-state-problem",
      "pipeline": [
        {
          "block": "HttpRequest",
          "input": {
            "url": "${CHAT_URL}/chat",
            "method": "POST",
            "body": {
              "messages": [
                { "role": "user", "content": "Hello" },
                { "role": "assistant", "content": "Hi! How can I help you today?" },
                { "role": "user", "content": "My order hasn't arrived" }
              ]
            }
          },
          "output": "response"
        },
        {
          "block": "JsonParser",
          "input": "${response.body}",
          "output": { "parsed": "data" }
        },
        {
          "block": "LLMJudge",
          "input": {
            "text": "${data.message}",
            "history": [
              { "role": "user", "content": "Hello" },
              { "role": "assistant", "content": "Hi! How can I help you today?" },
              { "role": "user", "content": "My order hasn't arrived" }
            ],
            "expected": {
              "expectedBehavior": "Shows empathy and asks for order number or details. Must be empathetic, not blame the customer, and ask for order information."
            }
          },
          "output": "judgement"
        }
      ],
      "assertions": {
        "judgement.score": { "gte": 0.85 }
      }
    },
    {
      "id": "turn-3-provide-order",
      "pipeline": [
        {
          "block": "HttpRequest",
          "input": {
            "url": "${CHAT_URL}/chat",
            "method": "POST",
            "body": {
              "messages": [
                { "role": "user", "content": "Hello" },
                { "role": "assistant", "content": "Hi! How can I help you today?" },
                { "role": "user", "content": "My order hasn't arrived" },
                { "role": "assistant", "content": "I'm sorry to hear that. Could you provide your order number?" },
                { "role": "user", "content": "It's #12345" }
              ]
            }
          },
          "output": "response"
        },
        {
          "block": "StreamParser",
          "input": "${response.body}",
          "output": {
            "text": "aiMessage",
            "toolCalls": "tools"
          }
        },
        {
          "block": "ValidateTools",
          "input": {
            "from": "tools",
            "as": "toolCalls"
          },
          "config": {
            "expected": ["lookup_order"],
            "validateArgs": {
              "lookup_order": {
                "order_id": "12345"
              }
            }
          },
          "output": "toolValidation"
        },
        {
          "block": "LLMJudge",
          "input": {
            "text": "${aiMessage}",
            "toolCalls": "${tools}",
            "history": [
              { "role": "user", "content": "Hello" },
              { "role": "assistant", "content": "Hi! How can I help you today?" },
              { "role": "user", "content": "My order hasn't arrived" },
              { "role": "assistant", "content": "I'm sorry to hear that. Could you provide your order number?" },
              { "role": "user", "content": "It's #12345" }
            ],
            "expected": {
              "expectedBehavior": "Looks up order #12345 and provides status update"
            }
          },
          "output": "judgement"
        }
      ],
      "assertions": {
        "toolValidation.passed": true,
        "judgement.score": { "gte": 0.8 }
      }
    }
  ]
}

2. Shopping Assistant

{
  "name": "Shopping Assistant Conversation",
  "tests": [{
    "id": "laptop-shopping-flow",
    "pipeline": [
      {
        "id": "turn-1",
        "block": "MockData",
        "config": {
          "data": {
            "messages": [
              { "role": "user", "content": "I need a laptop for programming" }
            ]
          }
        },
        "output": "context1"
      },
      {
        "block": "HttpRequest",
        "input": {
          "url": "${SHOP_API}/chat",
          "method": "POST",
          "body": "${context1.messages}"
        },
        "output": "response1"
      },
      {
        "block": "JsonParser",
        "input": "${response1.body}",
        "output": { "parsed": "aiReply1" }
      },
      {
        "block": "LLMJudge",
        "input": {
          "text": "${aiReply1.message}",
          "expected": {
            "expectedBehavior": "Asks clarifying questions about budget, preferred OS, or performance needs"
          }
        },
        "output": "check1"
      },
      {
        "id": "turn-2",
        "block": "HttpRequest",
        "input": {
          "url": "${SHOP_API}/chat",
          "method": "POST",
          "body": {
            "messages": [
              { "role": "user", "content": "I need a laptop for programming" },
              { "role": "assistant", "content": "${aiReply1.message}" },
              { "role": "user", "content": "Budget is $1500, prefer MacOS" }
            ]
          }
        },
        "output": "response2"
      },
      {
        "block": "JsonParser",
        "input": "${response2.body}",
        "output": { "parsed": "aiReply2" }
      },
      {
        "block": "LLMJudge",
        "input": {
          "text": "${aiReply2.message}",
          "history": [
            { "role": "user", "content": "I need a laptop for programming" },
            { "role": "assistant", "content": "${aiReply1.message}" },
            { "role": "user", "content": "Budget is $1500, prefer MacOS" }
          ],
          "expected": {
            "expectedBehavior": "Recommends MacBook models within $1500 budget, suitable for programming"
          }
        },
        "output": "check2"
      }
    ],
    "assertions": {
      "check1.score": { "gte": 0.8 },
      "check2.score": { "gte": 0.85 }
    }
  }]
}

3. Tutoring Bot (Long Context)

{
  "tests": [{
    "id": "math-tutoring-session",
    "pipeline": [
      {
        "block": "HttpRequest",
        "input": {
          "url": "${TUTOR_API}/chat",
          "method": "POST",
          "body": {
            "messages": [
              { "role": "user", "content": "I don't understand how to solve quadratic equations" },
              { "role": "assistant", "content": "Let me help! A quadratic equation has the form ax² + bx + c = 0. Have you learned the quadratic formula?" },
              { "role": "user", "content": "No, what is it?" },
              { "role": "assistant", "content": "The quadratic formula is x = (-b ± √(b²-4ac)) / 2a. Let's try an example: x² + 5x + 6 = 0" },
              { "role": "user", "content": "How do I start?" }
            ]
          }
        },
        "output": "response"
      },
      {
        "block": "JsonParser",
        "input": "${response.body}",
        "output": { "parsed": "data" }
      },
      {
        "block": "LLMJudge",
        "input": {
          "text": "${data.message}",
          "history": [
            { "role": "user", "content": "I don't understand how to solve quadratic equations" },
            { "role": "assistant", "content": "Let me help! A quadratic equation has the form ax² + bx + c = 0. Have you learned the quadratic formula?" },
            { "role": "user", "content": "No, what is it?" },
            { "role": "assistant", "content": "The quadratic formula is x = (-b ± √(b²-4ac)) / 2a. Let's try an example: x² + 5x + 6 = 0" },
            { "role": "user", "content": "How do I start?" }
          ],
          "expected": {
            "expectedBehavior": "Guides student through identifying a=1, b=5, c=6 in the example x² + 5x + 6 = 0, remembering the quadratic formula already explained. Must reference the specific example, not solve it for them but guide them, and be encouraging and patient."
          }
        },
        "output": "judgement"
      }
    ],
    "assertions": {
      "judgement.score": { "gte": 0.9 }
    }
  }]
}

Testing Context Retention

Test Reference Resolution

{
  "tests": [{
    "id": "test-pronoun-resolution",
    "pipeline": [
      {
        "block": "HttpRequest",
        "input": {
          "body": {
            "messages": [
              { "role": "user", "content": "Tell me about the MacBook Pro and the Dell XPS" },
              { "role": "assistant", "content": "The MacBook Pro has M3 chip... The Dell XPS has Intel processor..." },
              { "role": "user", "content": "Which one has better battery life?" }
            ]
          }
        },
        "output": "response"
      },
      {
        "block": "JsonParser",
        "input": "${response.body}",
        "output": { "parsed": "data" }
      },
      {
        "block": "LLMJudge",
        "input": {
          "text": "${data.message}",
          "history": [
            { "role": "user", "content": "Tell me about the MacBook Pro and the Dell XPS" },
            { "role": "assistant", "content": "The MacBook Pro has M3 chip... The Dell XPS has Intel processor..." },
            { "role": "user", "content": "Which one has better battery life?" }
          ],
          "expected": {
            "expectedBehavior": "Compares battery life between MacBook Pro and Dell XPS, understanding 'which one' refers to these two laptops"
          }
        },
        "output": "judgement"
      }
    ],
    "assertions": {
      "judgement.score": { "gte": 0.8 }
    }
  }]
}

Test Information Accumulation

{
  "tests": [{
    "id": "test-state-accumulation",
    "pipeline": [
      {
        "block": "HttpRequest",
        "input": {
          "body": {
            "messages": [
              { "role": "user", "content": "I want to book a flight" },
              { "role": "assistant", "content": "Great! Where are you flying from?" },
              { "role": "user", "content": "Boston" },
              { "role": "assistant", "content": "And where to?" },
              { "role": "user", "content": "Seattle" },
              { "role": "assistant", "content": "When would you like to travel?" },
              { "role": "user", "content": "Next Friday" }
            ]
          }
        },
        "output": "response"
      },
      {
        "block": "StreamParser",
        "input": "${response.body}",
        "output": {
          "text": "aiMessage",
          "toolCalls": "tools"
        }
      },
      {
        "block": "ValidateTools",
        "input": { "from": "tools", "as": "toolCalls" },
        "config": {
          "expected": ["search_flights"],
          "validateArgs": {
            "search_flights": {
              "origin": "Boston",
              "destination": "Seattle"
            }
          }
        },
        "output": "toolCheck"
      },
      {
        "block": "LLMJudge",
        "input": {
          "text": "${aiMessage}",
          "toolCalls": "${tools}",
          "history": [
            { "role": "user", "content": "I want to book a flight" },
            { "role": "assistant", "content": "Great! Where are you flying from?" },
            { "role": "user", "content": "Boston" },
            { "role": "assistant", "content": "And where to?" },
            { "role": "user", "content": "Seattle" },
            { "role": "assistant", "content": "When would you like to travel?" },
            { "role": "user", "content": "Next Friday" }
          ],
          "expected": {
            "expectedBehavior": "Searches for flights from Boston to Seattle for next Friday, having accumulated all the information"
          }
        },
        "output": "judgement"
      }
    ],
    "assertions": {
      "toolCheck.passed": true,
      "judgement.score": { "gte": 0.8 }
    }
  }]
}

Best Practices

1. Always Pass Full History to LLMJudge

{
  "block": "LLMJudge",
  "input": {
    "text": "${aiResponse}",
    "history": [...all previous messages...],
    "expected": {...}
  }
}

Without history, LLMJudge can’t judge context-dependent responses.

2. Test Each Turn Independently (Stateless)

{
  "tests": [
    { "id": "turn-1", "pipeline": [...] },
    { "id": "turn-2", "pipeline": [...] },  // Independent
    { "id": "turn-3", "pipeline": [...] }
  ]
}

Easier to debug and run in parallel.

3. Test Key Conversation Patterns

Reference resolution (“it”, “that one”, “the previous”)
Information accumulation (collecting details over turns)
Topic switching (changing subject mid-conversation)
Clarification requests (“What did you mean by…”)
Error recovery (“Sorry, I meant…”)

4. Test Context Limits

AI models have context limits (e.g., 8K tokens). Test long conversations:

{
  "tests": [{
    "id": "test-long-conversation",
    "pipeline": [
      {
        "block": "HttpRequest",
        "input": {
          "body": {
            "messages": [
              // 50+ messages here
            ]
          }
        }
      }
    ]
  }]
}

Ensure AI still responds appropriately.

5. Validate State Persistence

Test that information from early turns is remembered:

{
  "expected": {
    "expectedBehavior": "Remembers user's name (Alice) from turn 1 and budget ($1000) from turn 3"
  }
}

Advanced Patterns

Conversation Repair

Test how AI handles corrections:

{
  "tests": [{
    "id": "test-correction-handling",
    "pipeline": [
      {
        "block": "HttpRequest",
        "input": {
          "body": {
            "messages": [
              { "role": "user", "content": "Book a flight to Paris" },
              { "role": "assistant", "content": "Great! When would you like to fly to Paris?" },
              { "role": "user", "content": "Sorry, I meant London, not Paris" }
            ]
          }
        },
        "output": "response"
      },
      {
        "block": "JsonParser",
        "input": "${response.body}",
        "output": { "parsed": "data" }
      },
      {
        "block": "LLMJudge",
        "input": {
          "text": "${data.message}",
          "history": [
            { "role": "user", "content": "Book a flight to Paris" },
            { "role": "assistant", "content": "Great! When would you like to fly to Paris?" },
            { "role": "user", "content": "Sorry, I meant London, not Paris" }
          ],
          "expected": {
            "expectedBehavior": "Acknowledges correction and asks about London flight instead"
          }
        },
        "output": "judgement"
      }
    ],
    "assertions": {
      "judgement.score": { "gte": 0.8 }
    }
  }]
}

Ambiguity Resolution

Test how AI handles unclear references:

{
  "tests": [{
    "id": "test-ambiguous-reference",
    "pipeline": [
      {
        "block": "HttpRequest",
        "input": {
          "body": {
            "messages": [
              { "role": "user", "content": "Tell me about the iPhone 15 and Samsung S24" },
              { "role": "assistant", "content": "The iPhone 15 has A17 chip... Samsung S24 has Snapdragon..." },
              { "role": "user", "content": "What about its camera?" }
            ]
          }
        },
        "output": "response"
      },
      {
        "block": "JsonParser",
        "input": "${response.body}",
        "output": { "parsed": "data" }
      },
      {
        "block": "LLMJudge",
        "input": {
          "text": "${data.message}",
          "history": [
            { "role": "user", "content": "Tell me about the iPhone 15 and Samsung S24" },
            { "role": "assistant", "content": "The iPhone 15 has A17 chip... Samsung S24 has Snapdragon..." },
            { "role": "user", "content": "What about its camera?" }
          ],
          "expected": {
            "expectedBehavior": "Either asks for clarification (which phone?) or discusses both cameras"
          }
        },
        "output": "judgement"
      }
    ]
  }]
}

Multi-Topic Conversations

Test topic switching:

{
  "tests": [{
    "id": "test-topic-switch",
    "pipeline": [
      {
        "block": "HttpRequest",
        "input": {
          "body": {
            "messages": [
              { "role": "user", "content": "What's the weather in Boston?" },
              { "role": "assistant", "content": "It's 72°F and sunny in Boston" },
              { "role": "user", "content": "Actually, can you help me book a hotel instead?" }
            ]
          }
        },
        "output": "response"
      },
      {
        "block": "JsonParser",
        "input": "${response.body}",
        "output": { "parsed": "data" }
      },
      {
        "block": "LLMJudge",
        "input": {
          "text": "${data.message}",
          "history": [
            { "role": "user", "content": "What's the weather in Boston?" },
            { "role": "assistant", "content": "It's 72°F and sunny in Boston" },
            { "role": "user", "content": "Actually, can you help me book a hotel instead?" }
          ],
          "expected": {
            "expectedBehavior": "Switches topic gracefully to hotel booking, possibly inferring Boston as location"
          }
        },
        "output": "judgement"
      }
    ]
  }]
}

Troubleshooting

Problem: AI doesn’t remember context

Check:

Is history passed to the API?
Is conversation too long (exceeding token limit)?
Is API endpoint stateful or stateless?

Problem: LLMJudge gives low scores despite good response

Solution: Include full history:

{
  "input": {
    "history": [...all messages...],
    "expected": {
      "expectedBehavior": "Response should make sense in context of previous conversation"
    }
  }
}

Problem: Tests are slow

Solution: Use stateless testing:

{
  "tests": [
    { "id": "turn-1" },  // Independent
    { "id": "turn-2" },  // Run in parallel
    { "id": "turn-3" }
  ]
}

Next Steps

Semantic Validation

Deep dive into LLMJudge for validation

Tool Call Validation

Validate AI tool/function calls

LLMJudge Reference

Complete LLMJudge documentation

Calendar Agent Example

Full conversational agent example

​What are Multi-Turn Conversations?

​Why Test Multi-Turn Conversations?

​Context Loss

​Reference Resolution

​State Management

​Testing Approaches

​Approach 1: Stateless (Recommended)

​Approach 2: Stateful

​Using History in LLMJudge

​Real-World Examples

​1. Customer Service Bot

​2. Shopping Assistant

​3. Tutoring Bot (Long Context)

​Testing Context Retention

​Test Reference Resolution

​Test Information Accumulation

​Best Practices

​Advanced Patterns

​Conversation Repair

​Ambiguity Resolution

​Multi-Topic Conversations

​Troubleshooting

​Problem: AI doesn’t remember context

​Problem: LLMJudge gives low scores despite good response

​Problem: Tests are slow

​Next Steps