Strict Mode vs Flexible Mode

FlowKit supports two conversation modes that control how the LLM interacts with users.

Overview

Aspect	Flexible Mode	Strict Mode
LLM Freedom	High	Minimal
Message Control	LLM generates responses	Uses defined messages exactly
Conversation Feel	Natural, adaptive	Consistent, predictable
Best For	General assistance	Enterprise, regulated flows
LLM Requirements	Any quality	Better extraction needed

Flexible Mode (Default)

The LLM has freedom to respond naturally while following the flow structure.

Characteristics

• LLM generates responses in its own words
• Can add personality and contextual responses
• More natural conversation feel
• Adapts to user's tone and style

Example

const assistant = agent("Support Bot")
  // flexible mode is default, no need to set it
  .personality("friendly and helpful")
  .build();

const supportFlow = flow("support", assistant)
  .ask("issue", "What can I help you with today?", text(), "issue")
  .then("confirm")
  .ask("confirm", "Got it! Anything else?", yesNo(), "more")
  .when({ "yes": "issue", "no": "bye" })
  .say("bye", "Thanks for reaching out!")
  .done()
  .build();

Conversation Example

Bot: Hey there! What can I help you with today?
User: my internet is slow
Bot: Oh no, slow internet is frustrating! I've noted that down. 
     Is there anything else you'd like help with?
User: no that's it
Bot: Great! Thanks so much for reaching out. 
     We'll look into your slow internet issue. Take care!

---

Strict Mode

The LLM uses exact messages defined in the flow with minimal deviation.

Characteristics

• Uses exact messages from step definitions
• Consistent experience every time
• Better for compliance and quality control
• Easier to audit and test

Example

const assistant = agent("Verification Bot")
  .strict()  // Strict mode enabled
  .personality("Follow the verification process exactly")
  .build();

const verifyFlow = flow("verify", assistant)
  .ask("name", "Please enter your full name:", name(), "full_name")
  .then("dob")
  .ask("dob", "What is your date of birth? (MM/DD/YYYY)", custom("date"), "birth_date")
  .then("confirm")
  .ask("confirm", "Is this information correct: {{full_name}}, born {{birth_date}}?", yesNo(), "confirmed")
  .when({ "yes": "success", "no": "name" })
  .say("success", "Thank you. Your identity has been verified.")
  .done()
  .build();

Conversation Example

Bot: Please enter your full name:
User: john smith
Bot: What is your date of birth? (MM/DD/YYYY)
User: january 15 1990
Bot: Is this information correct: John Smith, born 01/15/1990?
User: yes
Bot: Thank you. Your identity has been verified.

---

When to Use Each

Use Flexible Mode When:

• Building general-purpose chatbots
• User experience is priority
• Conversation needs to feel natural
• You want the LLM to handle edge cases gracefully
• Brand voice allows variation

Use Strict Mode When:

• Building enterprise/regulated applications
• Consistency is critical
• Legal/compliance requirements exist
• You need predictable, testable responses
• Exact wording matters (financial, healthcare, legal)
• You want to control the experience precisely

---

Hybrid Approach

You can use different agents for different flows:

// Casual greeting flow - flexible
const greeterAgent = agent("Greeter")
  .personality("Be warm and friendly")
  .build();

// Verification flow - strict
const verifierAgent = agent("Verifier")
  .strict()
  .personality("Follow verification exactly")
  .build();

const greetingFlow = flow("greeting", greeterAgent)
  .ask("hello", "Hi! How's it going?", text(), "greeting")
  .then("transfer")
  .say("transfer", "Let me transfer you to verification...")
  .done()
  .build();

const verifyFlow = flow("verify", verifyAgent)
  // strict verification steps...
  .build();

// Use different engines for different flows
const greetingEngine = new FlowEngine(greetingFlow, { llm: adapter, storage });
const verifyEngine = new FlowEngine(verifyFlow, { llm: adapter, storage });

---

Technical Differences

System Prompt

Flexible mode system prompt allows LLM freedom:

You are [Agent Name]. [Instructions]
Current step: [step_id]
Ask: [message]
...respond naturally...

Strict mode system prompt enforces exact messages:

You are [Agent Name]. [Instructions]

IMPORTANT: Use EXACTLY the message defined for the current step.
Do not paraphrase or add to the message.

Current step: [step_id]
Say EXACTLY: "[message]"

Extraction

Both modes use JSON extraction, but strict mode is more demanding:

// Both modes extract the same way
// But strict mode requires better extraction accuracy
// because there's no room for follow-up clarification

---

LLM Model Considerations

Flexible Mode

Works well with most models:

• GPT-4o-mini (recommended)
• GPT-4o
• Llama 3
• Qwen
• Most 3B+ models

Strict Mode

Requires models with good instruction following and JSON extraction:

• GPT-5.2
• GPT-5 mini
• GPT-5 nano (recommended)
• GPT-4o
• GPT-4o-mini
• Claude 3
• Qwen 4B+
• Llama 3.2 3B (may struggle with JSON)

Tip: For strict mode with local models, use qwen3:4b or larger.

---

Migration Between Modes

Changing from flexible to strict (or vice versa) is simple:

// Before (flexible)
const assistant = agent("Bot").build();
  // mode defaults to "flexible"

// After (strict) - just add .strict()
const assistant = agent("Bot")
  .strict()
  .build();

No flow changes needed - only the agent configuration changes.