LLMs as Transitions: The Intent Trigger

Now we bring the two worlds together. If an FSM is the "Map," and the LLM is the "Driver," how does the driver tell the map where to go?

The answer is Intent Classification.

1. The Classifier Node

Instead of letting the LLM "Run a Tool," we let the LLM "Choose a State."

1. User: "I'd like to return this shirt."
2. LLM: Classifies the intent as INIT_RETURN.
3. FSM: Validates that a return is allowed for the current user. If yes, it moves to the RETURN_STARTED state.

---

2. Why this is 10x More Reliable

In a pure agent, the LLM might decide to:

• Call check_refund_policy.
• Call search_email.
• Call generate_label. ...in any order it wants.

In a hybrid system, the LLM is restricted to a specific list of intent categories.

---

3. Visualizing the Hand-off

graph LR
    Input[User: 'Cancel my sub'] --> Brain[LLM Classifier]
    Brain -->|Intent: CANCEL| FSM[State Machine]
    FSM -->|Current State: ACTIVE| NodeA[Process Cancellation]
    FSM -->|Current State: ALREADY_CANCELLED| NodeB[Explain Status]
    NodeA --> Result[Success Response]

---

4. Reducing Token Costs

Classifying an intent into 5 categories is much "Cheaper" than asking an agent to "Think" and "Plan."

• You can use a tiny model (Llama-3 8B or Phi-3) for the Classifier.
• You only use the big, expensive model for the Actual Content (writing the email).

---

5. Code Example: Triggering a transition

# 1. LLM output
intent = llm.call("Classify this message: 'I want to pay'. Categories: [LOGIN, PAY, LOGOUT]")

# 2. Logic check
if intent == "PAY":
    try:
        fsm_engine.trigger('pay') # This will fail if not logged in!
    except Exception:
        return "You must login before paying."

---

Key Takeaways

• The LLM acts as the dynamic interpreter of human language.
• The FSM acts as the deterministic guard of business rules.
• Intent Classification is the most robust way to trigger state changes.
• This pattern allows for smaller, faster models to handle the navigation.