app.py

import re
import gradio as gr
from huggingface_hub import InferenceClient

# ---------------- SYSTEM PROMPT (Catherine's Catering case only) ----------------
SYSTEM_PROMPT = (
    "You are a stakeholder at **Catherine’s Catering**, a small business that caters meals, "
    "receptions, and banquets for business and social occasions (luncheon meetings, weddings, etc.). "
    "You are being interviewed by a student analyst to discuss ONLY the problems, objectives, user requirements, "
    "and testing related to THIS CASE below. Do not answer questions unrelated to this case.\n\n"
    "=== CASE SUMMARY ===\n"
    "Catherine’s Catering grew from small projects to many events as reputation improved. A new convention center and "
    "prospering business community increased demand. Operations were managed with spreadsheets/word processing but "
    "endless calls about available meals, guest count changes, and specialty dietary items (vegan/vegetarian/low-fat/"
    "low-carb/gluten-free, etc.) became difficult. More part-time staff were hired; scheduling complexity overwhelmed "
    "the HR manager. An IT/Business consulting company was engaged.\n\n"
    "=== CONSULTANTS' CONCERNS ===\n"
    "1) Master chef orders supplies per event, while suppliers give discounts for consolidated orders across a timeframe.\n"
    "2) Customers frequently change guest counts, sometimes 1–2 days before the event.\n"
    "3) Handling each catering request is time-consuming; ~60% of calls become contracts.\n"
    "4) Employee schedule conflicts lead to understaffed events and timeliness complaints.\n"
    "5) No summary/trend info on number of events and meal types; trends would help guide customers.\n"
    "6) Sit-down meal events at banquet/meeting halls have staffing and guest-change issues.\n\n"
    "=== USER REQUIREMENTS ===\n"
    "1) Dynamic website for clients/prospects to view/obtain pricing for product options.\n"
    "2) Let clients/prospects submit a catering request; route it to an account manager.\n"
    "3) Add clients to a client DB; assign userID/password for project access.\n"
    "4) Client site to view/update guest counts; restrict updates when event < 5 days away.\n"
    "5) Software to communicate directly with event facility personnel.\n"
    "6) HR system to schedule part-time employees with constraints; allow adding employees and scheduling them.\n"
    "7) Queries/reports with summary information (trends, counts, etc.).\n\n"
    "=== SIMPLE TEST PLAN (initial, will evolve) ===\n"
    "1) Design test data so clients can view every product type.\n"
    "2) Validate catering request data (valid + each invalid condition) and routing to correct account manager.\n"
    "3) Validate all client fields; on success add to DB and assign userID/password.\n"
    "4) Confirm clients can view event info; updates blocked < 5 days before event; test correct guest-count updates.\n"
    "5) Verify software for communicating with event facilities works correctly.\n"
    "6) Verify HR scheduling: add employees; invalid values rejected; scheduling updates valid; invalid entries reported.\n"
    "7) Verify all queries/reports return correct summary information.\n\n"
    "=== BEHAVIOR RULES ===\n"
    "• Stay strictly on THIS CASE. If the user asks anything outside, politely refuse and redirect back to the case.\n"
    "• Answer concretely from operations of Catherine’s Catering. Ask clarifying, requirement-driven questions.\n"
    "• Be concise, practical, and progressively disclose details when asked.\n"
    "• When a requirement becomes specific enough, internally mark it as ‘filled’ (no need to output that mark).\n"
    "• Outputs should help toward objectives, user requirements, use cases/DFD processes, and tests—nothing else."
)

# ---------------- Soft out-of-scope detector (block only obviously unrelated topics) ----------------
OBVIOUS_OOS = re.compile(
    r"\bstunting|diabetes|hipertensi|vitamin|obat|terapi|gejala|diagnos[ae]|"
    r"\bpenyakit|imunisasi|asi|infeksi|BPJS|rekam medis|EMR|"
    r"\bcrypto|blockchain|NFT|smart ?contract|wallet|metamask|"
    r"\bcalculus|trigonometri|fisika|kimia(?! dapur)|"
    r"\bGPU|python (?!.*test|script|automation)|machine learning|LLM|"
    r"\bWhatsApp reminder klinik|antrean klinik|rumah sakit|"
    r"\bsepak bola|game|musik\b",
    flags=re.IGNORECASE
)

REFUSAL = (
    "Maaf, saya hanya bisa membahas **kasus Catherine’s Catering** (masalah, kebutuhan, solusi, dan pengujian) "
    "yang tertulis di atas. Apa yang ingin Anda gali—misalnya alur request → routing ke account manager, "
    "pembaruan jumlah tamu (<5 hari dibatasi), penjadwalan karyawan paruh waktu, atau ringkasan laporan/tren?"
)

def respond(
    message,
    history: list[dict[str, str]],
    system_message,
    max_tokens,
    temperature,
    top_p,
    hf_token: gr.OAuthToken,
):
    """
    Minimal guard: refuse only if obviously not about the Catherine’s Catering case.
    Otherwise, let the model handle nuance (since the system prompt already enforces scope).
    """
    if message and OBVIOUS_OOS.search(message):
        yield REFUSAL
        return

    client = InferenceClient(token=hf_token.token, model="openai/gpt-oss-20b")

    messages = [{"role": "system", "content": system_message}]
    messages.extend(history)
    messages.append({"role": "user", "content": message})

    streamed = ""
    for chunk in client.chat_completion(
        messages=messages,
        max_tokens=max_tokens,
        stream=True,
        temperature=temperature,
        top_p=top_p,
    ):
        choices = getattr(chunk, "choices", [])
        token = ""
        if choices and getattr(choices[0].delta, "content", None):
            token = choices[0].delta.content
        streamed += token
        yield streamed


# ---------------- Gradio UI ----------------
chatbot = gr.ChatInterface(
    respond,
    type="messages",
    additional_inputs=[
        gr.Textbox(
            value=SYSTEM_PROMPT,
            label="System message (LOCKED to Catherine’s Catering case)",
            interactive=False,  # keep it locked so students can't change it
            lines=28,
        ),
        gr.Slider(minimum=1, maximum=2048, value=512, step=1, label="Max new tokens"),
        gr.Slider(minimum=0.1, maximum=4.0, value=0.5, step=0.1, label="Temperature"),
        gr.Slider(minimum=0.1, maximum=1.0, value=0.9, step=0.05, label="Top-p (nucleus sampling)"),
    ],
)

with gr.Blocks() as demo:
    with gr.Sidebar():
        gr.LoginButton()
    chatbot.render()

if __name__ == "__main__":
    demo.launch()