code.py

"""Routing — classify the request first, then send it to a specialist.

One cheap classification call picks the lane; the chosen specialist prompt
handles the actual work. Adding a new lane means adding a dict entry, not
touching the ones that already work.

If the classifier isn't confident, we don't guess — we ask for clarification.
"""

from langchain_core.prompts import ChatPromptTemplate

from resources.agent import llm
from resources.helper import show_response

# The router. Kept deliberately small: its only job is to name a lane and rate
# its own confidence. Cheap prompt in, two words out.
route_prompt = ChatPromptTemplate.from_template(
    "Classify the customer message into exactly one category: "
    "tech, sales, or billing.\n"
    "Reply with the category, a space, then your confidence 0.0-1.0. "
    "Nothing else. Example: tech 0.9\n\nMessage: {message}"
)

# The specialists. Each one assumes it already has the right kind of request,
# which is exactly what lets its prompt be narrow and good at one thing.
SPECIALISTS = {
    "tech": "You are a technical support engineer. Diagnose the issue and give "
    "concrete troubleshooting steps.\n\nIssue: {message}",
    "sales": "You are a sales rep. Answer the pricing or plan question and "
    "suggest a next step.\n\nQuestion: {message}",
    "billing": "You are a billing specialist. Explain the charge or process the "
    "refund request clearly.\n\nRequest: {message}",
}

CONFIDENCE_FLOOR = 0.6  # below this we ask rather than guess


def route(message):
    """Classify, then dispatch. Returns the specialist's response message."""
    decision = llm.invoke(route_prompt.format(message=message))
    show_response(decision)

    # Parse "tech 0.9" defensively — the model is being asked for a rigid
    # format, but a router that crashes on odd output is worse than one that
    # falls back to asking a human.
    parts = decision.content.strip().split()
    lane = parts[0].lower().strip(".,") if parts else ""
    try:
        confidence = float(parts[1])
    except (IndexError, ValueError):
        confidence = 0.0

    if lane not in SPECIALISTS or confidence < CONFIDENCE_FLOOR:
        print(f"\n[router] lane={lane!r} confidence={confidence} -> too unsure, asking")
        return llm.invoke(
            "Ask the customer one short clarifying question about this "
            f"ambiguous message: {message}"
        )

    print(f"\n[router] lane={lane!r} confidence={confidence} -> dispatching")
    return llm.invoke(SPECIALISTS[lane].format(message=message))


if __name__ == "__main__":
    # Three messages, three lanes — same entry point, different specialists.
    for message in [
        "My deployment keeps failing with a 502 after the last update.",
        "What's the price difference between the team and enterprise plan?",
        "hey",  # deliberately vague: should trip the confidence floor
    ]:
        print(f"\n{'=' * 60}\nINCOMING: {message}\n{'=' * 60}")
        show_response(route(message))