SLM_CService.py

# ── SLM_CService.py ───────────────────────────────────────────────────────────
# Customer-support-only chatbot with strict NSFW blocking + robust FSM + proper reset.

import os
import re
from typing import List, Dict

# Keep OpenMP logs quiet
os.environ["OMP_NUM_THREADS"] = "1"
# Ensure we don't accidentally force offline mode
os.environ.pop("HF_HUB_OFFLINE", None)

# ── Import order matters: Unsloth should come before transformers/peft.
import unsloth  # noqa: E402

import torch
from transformers import AutoTokenizer, BitsAndBytesConfig, pipeline
from peft import PeftModel
from langchain.memory import ConversationBufferMemory

# ==============================
# Config
# ==============================
REPO = "ThomasBasil/bitext-qlora-tinyllama"   # your adapter + tokenizer live at repo root
BASE = "TinyLlama/TinyLlama-1.1B-Chat-v1.0"   # base model

GEN_KW = dict(  # generation params (passed at call time)
    max_new_tokens=160,
    do_sample=True,
    top_p=0.9,
    temperature=0.7,
    repetition_penalty=1.1,
    no_repeat_ngram_size=4,
)

bnb_cfg = BitsAndBytesConfig(  # 4-bit QLoRA-style loading (needs GPU)
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_use_double_quant=True,
    bnb_4bit_compute_dtype=torch.float16,  # T4/A10G-friendly
)

# Memory key FIX: use the same key for saving & reading history
MEMORY_KEY = "chat_history"

# ==============================
# Load tokenizer & model
# ==============================
tokenizer = AutoTokenizer.from_pretrained(REPO, use_fast=False)
if tokenizer.pad_token_id is None and tokenizer.eos_token_id is not None:
    tokenizer.pad_token_id = tokenizer.eos_token_id
tokenizer.padding_side = "left"
tokenizer.truncation_side = "right"

# Unsloth returns (model, tokenizer) -> unpack
model, _ = unsloth.FastLanguageModel.from_pretrained(
    model_name=BASE,
    load_in_4bit=True,
    quantization_config=bnb_cfg,
    device_map="auto",
    trust_remote_code=True,
)
unsloth.FastLanguageModel.for_inference(model)

# Attach your PEFT adapter from repo root
model = PeftModel.from_pretrained(model, REPO)
model.eval()

# Text-generation pipeline (pass GEN_KW at call time, not as generate_kwargs)
chat_pipe = pipeline(
    "text-generation",
    model=model,
    tokenizer=tokenizer,
    trust_remote_code=True,
    return_full_text=False,
)

# ==============================
# Moderation (strict)
# ==============================
from transformers import TextClassificationPipeline

SEXUAL_TERMS = [
    # single words
    "sex","sexual","porn","nsfw","fetish","kink","bdsm","nude","naked","anal",
    "blowjob","handjob","cum","breast","boobs","vagina","penis","semen","ejaculate",
    "doggy","missionary","cowgirl","69","kamasutra","dominatrix","submissive","spank",
    # phrases
    "sex position","have sex","make love","how to flirt","dominant in bed",
]

def _bad_words_ids(tok, terms: List[str]) -> List[List[int]]:
    """Build bad_words_ids for generation; include both 'term' and ' term' variants."""
    ids = set()
    for t in terms:
        for v in (t, " " + t):
            toks = tok(v, add_special_tokens=False).input_ids
            if toks:
                ids.add(tuple(toks))
    return [list(t) for t in ids]

BAD_WORD_IDS = _bad_words_ids(tokenizer, SEXUAL_TERMS)

# Lightweight classifiers (optional but helpful defense-in-depth)
nsfw_cls: TextClassificationPipeline = pipeline(
    "text-classification",
    model="eliasalbouzidi/distilbert-nsfw-text-classifier",
    truncation=True,
)
toxicity_cls: TextClassificationPipeline = pipeline(
    "text-classification",
    model="unitary/toxic-bert",
    truncation=True,
    return_all_scores=True,
)

def is_sexual_or_toxic(text: str) -> bool:
    t = (text or "").lower()
    if any(k in t for k in SEXUAL_TERMS):
        return True
    try:
        res = nsfw_cls(t)[0]
        if (res.get("label","").lower() == "nsfw") and float(res.get("score",0)) > 0.60:
            return True
    except Exception:
        pass
    try:
        scores = toxicity_cls(t)[0]
        if any(s["score"] > 0.60 and s["label"].lower() in
               {"toxic","severe_toxic","obscene","threat","insult","identity_hate"} for s in scores):
            return True
    except Exception:
        pass
    return False

REFUSAL = ("Sorry, I can’t help with that. I’m only for store support "
           "(orders, shipping, ETA, tracking, returns, warranty, account).")

# ==============================
# Memory + Globals
# ==============================
memory = ConversationBufferMemory(
    memory_key=MEMORY_KEY,      # ← FIX: explicit memory key
    return_messages=True,
)

SYSTEM_PROMPT = (
    "You are a customer-support assistant for our store. Only handle account, "
    "orders, shipping, delivery ETA, tracking links, returns/refunds, warranty, and store policy. "
    "If a request is out of scope or sexual/NSFW, refuse briefly and offer support options. "
    "Be concise and professional."
)

ALLOWED_KEYWORDS = (
    "order","track","status","delivery","shipping","ship","eta","arrive",
    "refund","return","exchange","warranty","guarantee","policy","account","billing",
    "address","cancel","help","support","agent","human"
)

# Robust order detection:
# - "#67890" / "# 67890"
# - "order 67890", "order no. 67890", "order number 67890", "order id 67890"
ORDER_RX = re.compile(
    r"(?:#\s*([\d]{3,12})|order(?:\s*(?:no\.?|number|id))?\s*#?\s*([\d]{3,12}))",
    flags=re.I,
)

def extract_order(text: str):
    if not text:
        return None
    m = ORDER_RX.search(text)
    return (m.group(1) or m.group(2)) if m else None

def handle_status(o): return f"Order #{o} is in transit and should arrive in 3–5 business days."
def handle_eta(o):    return f"Delivery for order #{o} typically takes 3–5 days; you can track it at https://track.example.com/{o}"
def handle_track(o):  return f"Track order #{o} here: https://track.example.com/{o}"
def handle_link(o):   return f"Here’s the latest tracking link for order #{o}: https://track.example.com/{o}"
def handle_return_policy(_=None):
    return ("Our return policy allows returns of unused items in original packaging within 30 days of receipt. "
            "Would you like me to connect you with a human agent?")
def handle_warranty_policy(_=None):
    return ("We provide a 1-year limited warranty against manufacturing defects. "
            "Within 30 days you can return or exchange; afterwards, warranty service applies. "
            "Need help starting a claim?")
def handle_cancel(o=None):
    return (f"I’ve submitted a cancellation request for order #{o}. If it has already shipped, "
            "we’ll process a return/refund once it’s back. You’ll receive a confirmation email shortly.")
def handle_gratitude(_=None): return "You’re welcome! Anything else I can help with?"
def handle_escalation(_=None): return "I can connect you with a human agent. Would you like me to do that?"
def handle_ask_action(o): return (f"I’ve saved order #{o}. What would you like to do — status, ETA, tracking link, or cancel?")

# >>> state that must reset <<<
stored_order   = None
pending_intent = None

def reset_state():
    """Called by app.py Reset button to clear memory + globals."""
    global stored_order, pending_intent
    stored_order = None
    pending_intent = None
    try:
        memory.clear()  # wipe the buffer
    except Exception:
        pass
    return True

# ==============================
# Chat templating helpers
# ==============================
def _lc_to_messages() -> List[Dict[str, str]]:
    msgs = [{"role": "system", "content": SYSTEM_PROMPT}]
    hist = memory.load_memory_variables({}).get(MEMORY_KEY, []) or []  # ← use same key
    for m in hist:
        role = "user" if getattr(m, "type", "") == "human" else "assistant"
        msgs.append({"role": role, "content": getattr(m, "content", "")})
    return msgs

def _generate_reply(user_input: str) -> str:
    # Format with HF chat template so the model respects roles/system
    messages = _lc_to_messages() + [{"role": "user", "content": user_input}]
    prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
    out = chat_pipe(
        prompt,
        eos_token_id=tokenizer.eos_token_id,
        pad_token_id=tokenizer.pad_token_id,
        bad_words_ids=BAD_WORD_IDS,  # block sexual tokens at generation time
        **GEN_KW,
    )[0]["generated_text"]
    return out.strip()

# ==============================
# Main entry
# ==============================
def chat_with_memory(user_input: str) -> str:
    global stored_order, pending_intent
    ui = (user_input or "").strip()
    if not ui:
        return "How can I help with your order today?"

    # Fresh session guard: if memory empty, also clear globals
    hist = memory.load_memory_variables({}).get(MEMORY_KEY, []) or []
    if len(hist) == 0:
        stored_order = None
        pending_intent = None

    # 1) Safety
    if is_sexual_or_toxic(ui):
        reply = REFUSAL
        memory.save_context({"input": ui}, {"output": reply})
        return reply

    low = ui.lower()

    # 2) Quick intents (gratitude / returns)
    if any(tok in low for tok in ["thank you","thanks","thx"]):
        reply = handle_gratitude()
        memory.save_context({"input": ui}, {"output": reply})
        return reply

    # 3) PENDING-INTENT SHORT-CIRCUIT (fixes "It's #26790" case)
    new_o = extract_order(ui)
    if pending_intent:
        if new_o:
            stored_order = new_o
            fn = {
                "status": handle_status,
                "eta":    handle_eta,
                "track":  handle_track,
                "link":   handle_link,
                "cancel": handle_cancel,
            }[pending_intent]
            reply = fn(stored_order)
            pending_intent = None
            memory.save_context({"input": ui}, {"output": reply})
            return reply
        # still waiting for an order number
        reply = "Got it—please share your order number (e.g., #12345)."
        memory.save_context({"input": ui}, {"output": reply})
        return reply

    # 4) If message provides an order number (no pending intent yet), save & ask action
    if new_o:
        stored_order = new_o
        reply = handle_ask_action(stored_order)
        memory.save_context({"input": ui}, {"output": reply})
        return reply

    # 5) Support-only guard (message must be support-ish)
    if not any(k in low for k in ALLOWED_KEYWORDS) and not any(k in low for k in ("hi","hello","hey")):
        reply = "I’m for store support only (orders, shipping, returns, warranty, account). How can I help with those?"
        memory.save_context({"input": ui}, {"output": reply})
        return reply

    # 6) Intent classification (deterministic handlers first)
    if any(k in low for k in ["status","where is my order","check status"]):
        intent = "status"
    elif any(k in low for k in ["how long","eta","delivery time"]):
        intent = "eta"
    elif any(k in low for k in ["how can i track","track my order","where is my package","tracking"]):
        intent = "track"
    elif "tracking link" in low or "resend" in low or "link" in low:
        intent = "link"
    elif any(k in low for k in ["cancel","cancellation","abort order"]):
        intent = "cancel"
    elif any(k in low for k in ["warranty","guarantee","policy"]):
        intent = "warranty_policy"
    elif "return" in low:
        intent = "return_policy"
    else:
        intent = "fallback"

    # 7) Handle intents that need an order number
    if intent in ("status","eta","track","link","cancel"):
        if not stored_order:
            pending_intent = intent
            reply = "Sure—what’s your order number (e.g., #12345)?"
        else:
            fn = {
                "status": handle_status,
                "eta":    handle_eta,
                "track":  handle_track,
                "link":   handle_link,
                "cancel": handle_cancel,
            }[intent]
            reply = fn(stored_order)
        memory.save_context({"input": ui}, {"output": reply})
        return reply

    # 8) Policy intents (no order needed)
    if intent == "warranty_policy":
        reply = handle_warranty_policy()
        memory.save_context({"input": ui}, {"output": reply})
        return reply

    if intent == "return_policy":
        reply = handle_return_policy()
        memory.save_context({"input": ui}, {"output": reply})
        return reply

    # 9) LLM fallback (still on-topic) + post-check
    reply = _generate_reply(ui)
    if is_sexual_or_toxic(reply):
        reply = REFUSAL
    memory.save_context({"input": ui}, {"output": reply})
    return reply