chat.py

import os
import re
import json
import time
import random
from datetime import datetime
from pathlib import Path
from collections import deque
from typing import Optional, Dict, Any, List, Tuple

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM, TextIteratorStreamer
from threading import Thread
from rich.markup import escape as rich_escape
from rich.panel import Panel
from rich.prompt import Prompt, Confirm
from rich.table import Table
from rich import box

from config import MODEL_DIR
from models import ModelManager
from utils import console, Theme, error_logger, debug_logger, get_device


# ---------------------------------------------------------------------------
# Command aliases (typo-tolerant)
# ---------------------------------------------------------------------------
_COMMAND_ALIASES = {
    "xlear": "clear", "claer": "clear", "clea": "clear",
    "exi": "exit", "ext": "exit", "quitt": "exit",
    "hlp": "help", "hep": "help",
    "hist": "history", "his": "history",
    "rg": "regen", "regen": "regen",
    "u": "undo",
    "s": "stats", "st": "stats",
    "e": "exit",
}

_DEFAULT_PERSONA = "Anda adalah asisten AI yang ramah dan informatif."

# Boilerplate umum dari model yang wajib dipotong
_BOILERPLATE_MARKERS = [
    "semoga membantu",
    "semoga informasinya jelas",
    "saya siap membantu lebih lanjut",
    "ada lagi yang ingin ditanyakan",
    "ada lagi yang bisa dibantu",
    "kalau ada pertanyaan lain",
    "apakah ada yang ingin didiskusikan",
    "jangan ragu untuk bertanya",
    "terima kasih sudah bertanya",
    "semoga harimu menyenangkan",
    "bisa ditanyakan lebih detail",
    "saya di sini untuk membantu",
    "mari jaga percakapan",
    "semoga kita bisa bicara",
    "saya mohon maaf atas ketidaknyamanan",
    "mari fokus pada solusi",
    "silakan ajukan pertanyaan yang sopan",
]

# Frasa sampah yang menandakan output ngelantur (typo atau halusinasi)
_JUNK_PHRASES = [
    "seminta", "semohon", "semiap", "sembuh", "semangat membantu",
    "saya di rumah", "hati-hati di jalan", "jangan lupa pemanasan",
    "jangan lupa pendinginan", "saya siap belajar",
    "saya catat", "hari baik", "di sana untuk membantu",
    "baik bicara lebih baik", "baik yang baik",
    "bicara lebih positif", "mari berdiskusi",
    "silakan bicara dengan tenang", "mari siap membantu",
]

# Kata kasar user yang sering bikin model output aneh
_RUDE_WORDS = {
    "anjing", "bangsat", "kontol", "memek", "goblok", "tolol", "idiot",
    "bego", "bodoh", "kampret", "sialan", "bajingan", "keparat",
}

# Fallback responses
_FALLBACKS = {
    "greeting": "Halo! Ada yang bisa saya bantu?",
    "howareyou": "Saya baik, terima kasih. Ada yang bisa dibantu?",
    "thanks": "Sama-sama! Senang bisa membantu.",
    "bye": "Sampai jumpa! Senang bisa membantu.",
    "rude": "Yuk, kita jaga percakapan tetap sopan ya. Ada yang bisa saya bantu?",
    "unknown": "Maaf, saya belum bisa menjawab itu dengan baik. Bisa dicoba dengan pertanyaan lain?",
    "short": "Bisa dijelaskan sedikit lebih detail?",
}

# Keyword untuk deteksi intent sederhana
_GREETINGS = {"halo", "hai", "hi", "hello", "woi", "hei", "assalamualaikum", "pagi", "siang", "sore", "malam"}
_THANKS = {"makasih", "terima kasih", "thanks", "thank you", "thx", "tq"}
_BYES = {"bye", "dadah", "sampai jumpa", "selamat tinggal", "goodbye"}

_SESSION_DIR = Path("/content/data/chat_sessions")


class EnhancedChatModule:
    def __init__(self):
        self.model_manager = ModelManager()
        self.device = get_device()
        self.tokenizer = None
        self.model = None

        self.conversation_history = deque(maxlen=50)
        self.max_context_tokens = 768
        self.system_prompt = _DEFAULT_PERSONA
        self.is_peft_model = False

        # Konfigurasi generation yang lebih ketat untuk model kecil
        self.gen_config = {
            "temperature": 0.75,
            "top_p": 0.9,
            "top_k": 40,
            "max_new_tokens": 60,           # pendek = lebih aman
            "repetition_penalty": 1.3,      # lebih tinggi dari sebelumnya
            "no_repeat_ngram_size": 4,      # lebih agresif
            "streaming": False,

            # Best-of-N sampling
            "num_attempts": 3,
            "min_acceptable_score": 0.45,
        }

        self.performance_stats = {
            "total_generations": 0,
            "avg_time_ms": 0.0,
            "total_tokens_generated": 0,
            "total_tokens_processed": 0,
            "max_response_time": 0.0,
            "min_response_time": float('inf'),
            "last_response": "",
            "last_user_input": "",
            "fallbacks_triggered": 0,
            "regenerations": 0,
        }

        _SESSION_DIR.mkdir(parents=True, exist_ok=True)

    # ==================================================================
    # Main loop
    # ==================================================================
    def run(self) -> None:
        current_model = self.model_manager.get_current_model()

        if not current_model:
            console.print(Theme.error(" Belum ada model. Training dulu!"))
            return

        model_path = os.path.join(MODEL_DIR, current_model)
        if not os.path.exists(model_path):
            console.print(Theme.error(" Model tidak ditemukan!"))
            return

        console.print(Theme.info(f"Memuat model: {current_model}..."))
        try:
            self._load_model(model_path)
            self.model.eval()
            console.print(Theme.success(" Model siap!"))
        except Exception as e:
            console.print(Theme.error(f"Gagal memuat model: {e}"))
            error_logger.error(f"Model loading error: {e}")
            return

        # Sync config dari ModelManager (temperature dll)
        self._sync_gen_config_from_manager()

        console.print(Panel(
            "[cyan] ENHANCED CHAT v2.2[/cyan]\n"
            "[white]Ketik 'help' untuk bantuan, 'exit' keluar[/white]\n"
            f"[dim]Model: {current_model} | Device: {self.device.upper()} | "
            f"Best-of-{self.gen_config['num_attempts']} | "
            f"Streaming: {'ON' if self.gen_config['streaming'] else 'OFF'}[/dim]",
            width=72, style="cyan"
        ))

        while True:
            try:
                user_input = Prompt.ask("[cyan]You").strip()
                if not user_input:
                    continue

                lower = user_input.lower().strip()
                if lower in _COMMAND_ALIASES:
                    lower = _COMMAND_ALIASES[lower]

                if self._handle_command(lower, user_input):
                    continue

                self._generate_and_display(user_input)

            except KeyboardInterrupt:
                console.print("\n" + Theme.warning("Chat dihentikan"))
                break
            except Exception as e:
                console.print(Theme.error(f"Error: {rich_escape(str(e))}"))
                error_logger.error(f"Chat error: {e}")
                if "CUDA" in str(e) or "out of memory" in str(e).lower():
                    try:
                        self.model = self.model.to('cpu')
                        torch.cuda.empty_cache()
                        self.device = 'cpu'
                        console.print(Theme.success(" Moved to CPU"))
                    except Exception:
                        break

    # ==================================================================
    # Command handler
    # ==================================================================
    def _handle_command(self, lower: str, raw: str) -> bool:
        if lower in ('exit', 'q'):
            console.print(Theme.success("Terima kasih! "))
            raise SystemExit

        if lower == 'clear':
            self.conversation_history.clear()
            console.print(Theme.success(" Riwayat dihapus"))
            return True

        if lower in ('history', 'hist'):
            self._show_history()
            return True

        if lower == 'stats':
            self._show_stats()
            return True

        if lower == 'help':
            self._show_help()
            return True

        if lower in ('undo', 'u'):
            if self.conversation_history:
                removed = self.conversation_history.pop()
                console.print(Theme.success(f"↩ Undo: hapus '{removed['user'][:40]}...'"))
            else:
                console.print(Theme.warning("Belum ada riwayat"))
            return True

        if lower in ('regen', 'rg'):
            last_user = self.performance_stats.get("last_user_input")
            if not last_user:
                console.print(Theme.warning("Belum ada percakapan"))
                return True
            if self.conversation_history:
                self.conversation_history.pop()
            self.performance_stats["regenerations"] += 1
            console.print(Theme.info(" Regenerate..."))
            self._generate_and_display(last_user)
            return True

        if lower == 'save':
            self._save_session()
            return True

        if lower == 'sessions':
            self._list_sessions()
            return True

        if lower.startswith('load'):
            parts = raw.split(maxsplit=1)
            if len(parts) < 2:
                self._list_sessions()
                return True
            try:
                idx = int(parts[1])
                self._load_session(idx)
            except ValueError:
                console.print(Theme.error("Format: load <nomor>"))
            return True

        if lower == 'export':
            self._export_chat()
            return True

        if lower.startswith('stream '):
            val = lower.split(maxsplit=1)[1]
            if val in ("on", "true", "1", "yes"):
                self.gen_config["streaming"] = True
                console.print(Theme.success(" Streaming: ON"))
            elif val in ("off", "false", "0", "no"):
                self.gen_config["streaming"] = False
                console.print(Theme.success(" Streaming: OFF"))
            else:
                console.print(Theme.error("Format: stream on|off"))
            return True

        if lower.startswith('temp '):
            try:
                val = float(lower.split(maxsplit=1)[1])
                if 0.01 <= val <= 2.0:
                    self.gen_config["temperature"] = val
                    console.print(Theme.success(f" Temperature → {val}"))
                else:
                    console.print(Theme.warning("Nilai harus 0.01-2.0"))
            except (ValueError, IndexError):
                console.print(Theme.error("Format: temp <0.01-2.0>"))
            return True

        if lower.startswith('topp '):
            try:
                val = float(lower.split(maxsplit=1)[1])
                if 0.0 <= val <= 1.0:
                    self.gen_config["top_p"] = val
                    console.print(Theme.success(f" Top-p → {val}"))
                else:
                    console.print(Theme.warning("Nilai harus 0.0-1.0"))
            except (ValueError, IndexError):
                console.print(Theme.error("Format: topp <0.0-1.0>"))
            return True

        if lower.startswith('maxtok '):
            try:
                val = int(lower.split(maxsplit=1)[1])
                if 10 <= val <= 500:
                    self.gen_config["max_new_tokens"] = val
                    console.print(Theme.success(f" Max new tokens → {val}"))
                else:
                    console.print(Theme.warning("Nilai harus 10-500"))
            except (ValueError, IndexError):
                console.print(Theme.error("Format: maxtok <10-500>"))
            return True

        if lower.startswith('attempts '):
            try:
                val = int(lower.split(maxsplit=1)[1])
                if 1 <= val <= 10:
                    self.gen_config["num_attempts"] = val
                    console.print(Theme.success(f" Best-of-N → {val}"))
                else:
                    console.print(Theme.warning("Nilai harus 1-10"))
            except (ValueError, IndexError):
                console.print(Theme.error("Format: attempts <1-10>"))
            return True

        if lower.startswith('persona '):
            new_persona = raw[8:].strip()
            if new_persona.lower() == "reset":
                self.system_prompt = _DEFAULT_PERSONA
                console.print(Theme.success(" Persona di-reset ke default"))
            elif new_persona:
                self.system_prompt = new_persona
                console.print(Theme.success(f" Persona di-set ({len(new_persona)} chars)"))
            return True

        if lower == 'persona':
            console.print(Panel(
                f"[cyan]Current persona:[/cyan]\n{rich_escape(self.system_prompt)}",
                title="PERSONA", style="cyan"
            ))
            return True

        return False

    # ==================================================================
    # Generate & display (BEST-OF-N)
    # ==================================================================
    def _generate_and_display(self, user_input: str) -> None:
        start = time.perf_counter()

        # ----- Fast path: fallback untuk greeting / thanks / bye / rude -----
        fast = self._fast_intent_response(user_input)
        if fast is not None:
            response = fast
        elif self.gen_config.get("streaming", False):
            response = self._generate_streaming(user_input)
        else:
            # Best-of-N
            response = self._generate_best_of_n(user_input)

        elapsed_ms = (time.perf_counter() - start) * 1000

        if not self.gen_config.get("streaming", False):
            console.print(f"[magenta]AI[/]: {rich_escape(response)}")
            console.print(Theme.dim(f" {elapsed_ms:.0f}ms | {len(response.split())} words"))
            console.print()

        self._update_stats(elapsed_ms, response)
        self.performance_stats["last_user_input"] = user_input
        self.performance_stats["last_response"] = response

        self.conversation_history.append({
            'user': user_input,
            'ai': response,
            'timestamp': time.time(),
        })

    # ==================================================================
    # Fast intent — deteksi sederhana tanpa model
    # ==================================================================
    def _fast_intent_response(self, user_input: str) -> Optional[str]:
        """Kalau intent jelas, jawab langsung. Model tidak dipanggil."""
        lower = user_input.lower().strip()
        words = set(re.findall(r'\w+', lower))

        # Kasar → jangan kasih ke model
        if words & _RUDE_WORDS:
            return _FALLBACKS["rude"]

        # Very short input (< 3 chars)
        if len(lower) < 3:
            return _FALLBACKS["short"]

        # Greeting
        if len(words) <= 3 and (words & _GREETINGS):
            return _FALLBACKS["greeting"]

        # Thanks
        if words & _THANKS:
            return _FALLBACKS["thanks"]

        # Bye
        if any(b in lower for b in _BYES):
            return _FALLBACKS["bye"]

        # "Apa kabar"
        if "kabar" in lower and len(words) <= 4:
            return _FALLBACKS["howareyou"]

        return None

    # ==================================================================
    # Best-of-N generation
    # ==================================================================
    def _generate_best_of_n(self, user_input: str) -> str:
        """Generate N kandidat, pilih skor tertinggi."""
        n = max(1, self.gen_config.get("num_attempts", 3))

        candidates: List[Tuple[float, str]] = []
        context = self._build_context(user_input)

        for i in range(n):
            try:
                raw = self._generate_raw(context, attempt_index=i)
                cleaned = self._clean_response(raw, user_input)
                score = self._score_response(cleaned, user_input)
                candidates.append((score, cleaned))
            except Exception as e:
                debug_logger.debug(f"Attempt {i} failed: {e}")
                continue

        if not candidates:
            self.performance_stats["fallbacks_triggered"] += 1
            return _FALLBACKS["unknown"]

        candidates.sort(key=lambda x: x[0], reverse=True)
        best_score, best_response = candidates[0]

        min_score = self.gen_config.get("min_acceptable_score", 0.4)
        if best_score < min_score:
            self.performance_stats["fallbacks_triggered"] += 1
            return _FALLBACKS["unknown"]

        return best_response

    def _generate_raw(self, context: str, attempt_index: int = 0) -> str:
        """Generate sekali. Return raw string dari model."""
        inputs = self.tokenizer.encode(
            context,
            return_tensors='pt',
            truncation=True,
            max_length=self.max_context_tokens,
        ).to(self.device)
        input_length = inputs.shape[1]

        model_max_len = getattr(self.model.config, 'max_position_embeddings', 1024)
        max_new_tokens = min(
            self.gen_config["max_new_tokens"],
            model_max_len - input_length - 10,
        )
        if max_new_tokens <= 0:
            keep_tokens = model_max_len - 50
            inputs = inputs[:, -keep_tokens:]
            input_length = inputs.shape[1]
            max_new_tokens = 30

        # Temperature bervariasi antar attempt → diversity
        base_temp = self.gen_config["temperature"]
        temp_var = base_temp + (attempt_index - 1) * 0.05
        temp_var = max(0.5, min(temp_var, 1.1))

        kwargs = {
            "max_new_tokens": max(1, max_new_tokens),
            "temperature": temp_var,
            "top_p": self.gen_config["top_p"],
            "top_k": self.gen_config["top_k"],
            "do_sample": True,
            "pad_token_id": self.tokenizer.eos_token_id,
            "eos_token_id": self.tokenizer.eos_token_id,
            "repetition_penalty": self.gen_config["repetition_penalty"],
            "no_repeat_ngram_size": self.gen_config["no_repeat_ngram_size"],
            "use_cache": True,
        }

        with torch.no_grad():
            outputs = self.model.generate(inputs, **kwargs)

        generated_tokens = outputs[0, input_length:]
        return self.tokenizer.decode(generated_tokens, skip_special_tokens=True)

    # ==================================================================
    # Response scoring
    # ==================================================================
    def _score_response(self, response: str, user_input: str) -> float:
        """
        Skor 0.0 - 1.0. Semakin tinggi semakin baik.
        Penalty untuk: junk phrases, repetitive, boilerplate, terlalu pendek.
        """
        if not response or not response.strip():
            return 0.0

        score = 1.0
        words = response.lower().split()
        word_count = len(words)

        # 1. Panjang wajar: 3 - 40 kata ideal
        if word_count < 3:
            score -= 0.5
        elif word_count < 5:
            score -= 0.2
        elif word_count > 40:
            score -= 0.3
        elif word_count > 60:
            score -= 0.6

        # 2. Uniqueness: kalau banyak kata diulang = jelek
        if word_count >= 5:
            unique_ratio = len(set(words)) / word_count
            if unique_ratio < 0.5:
                score -= 0.5
            elif unique_ratio < 0.65:
                score -= 0.2

        # 3. Junk phrases
        lower = response.lower()
        junk_count = sum(1 for p in _JUNK_PHRASES if p in lower)
        score -= junk_count * 0.25

        # 4. Boilerplate
        boilerplate_count = sum(1 for m in _BOILERPLATE_MARKERS if m in lower)
        score -= boilerplate_count * 0.15

        # 5. Terlalu banyak tanda tanya
        q_count = response.count("?")
        if q_count >= 3:
            score -= 0.3
        elif q_count >= 5:
            score -= 0.6

        # 6. Terlalu banyak koma = kalimat tidak terstruktur
        comma_count = response.count(",")
        if word_count > 0 and comma_count / max(word_count, 1) > 0.15:
            score -= 0.15

        # 7. Bonus: ada kata kunci user di response (relevan)
        user_words = set(w.lower() for w in re.findall(r'\w+', user_input) if len(w) > 3)
        if user_words:
            overlap = len(user_words & set(words))
            if overlap > 0:
                score += min(0.2, overlap * 0.05)

        # 8. Bonus: respon pendek 5-25 kata = ideal
        if 5 <= word_count <= 25 and not any(p in lower for p in _JUNK_PHRASES):
            score += 0.1

        return max(0.0, min(1.0, score))

    # ==================================================================
    # Response cleaning (agresif)
    # ==================================================================
    def _clean_response(self, response: str, user_input: str) -> str:
        if not response:
            return ""

        # 1. Buang user_input yang ikut ke-generate
        if user_input and len(user_input) > 3:
            response = response.replace(user_input, '').strip()

        # 2. Buang markup sisa
        response = re.sub(r'^(AI:|User:)', '', response).strip()
        for marker in ['User:', 'AI:', '<|endoftext|>', '<|end|>']:
            pos = response.find(marker)
            if pos != -1:
                response = response[:pos].strip()
        response = (
            response.replace('<|endoftext|>', '')
            .replace('<|end|>', '')
            .strip()
        )

        # 3. Rapikan whitespace
        response = ' '.join(response.split())

        # 4. Fix typo umum yang konsisten muncul
        typo_fixes = {
            "Seminta": "Semoga",
            "Semohon": "Semoga",
            "semiap": "siap",
        }
        for typo, fix in typo_fixes.items():
            response = response.replace(typo, fix)

        # 5. Cutoff di junk phrase pertama
        lower = response.lower()
        cutoff = len(response)
        for phrase in _JUNK_PHRASES:
            pos = lower.find(phrase)
            if pos != -1 and pos < cutoff:
                cutoff = pos
        if cutoff < len(response):
            response = response[:cutoff].rstrip(" ,;.")

        # 6. Cutoff di boilerplate pertama (setelah 25% awal)
        lower = response.lower()
        cutoff = len(response)
        min_pos = int(len(response) * 0.25)
        for marker in _BOILERPLATE_MARKERS:
            pos = lower.find(marker, min_pos)
            if pos != -1 and pos < cutoff:
                cutoff = pos
        if cutoff < len(response):
            response = response[:cutoff].rstrip(" ,;.")

        # 7. Fix punctuation aneh
        response = re.sub(r'[;:,]+\.', '.', response)
        response = re.sub(r'\.\.+', '.', response)
        response = re.sub(r'[,]{2,}', ',', response)
        response = re.sub(r'[;]{2,}', ';', response)
        response = re.sub(r'\s+([.,!?;:])', r'\1', response)
        response = re.sub(r'([.,!?;:])([^\s\d])', r'\1 \2', response)
        response = re.sub(r'\s+\.', '.', response)

        # 8. Batas kalimat: max 2
        sentences = re.split(r'(?<=[.!?])\s+', response)
        if len(sentences) > 2:
            response = ' '.join(sentences[:2]).strip()
            if response and response[-1] not in ".!?":
                response += "."

        # 9. Hard cap karakter
        if len(response) > 250:
            response = response[:250].rsplit(' ', 1)[0] + "..."

        return response.strip()

    # ==================================================================
    # Streaming (tetap ada)
    # ==================================================================
    def _generate_streaming(self, user_input: str) -> str:
        if not self.model or not self.tokenizer:
            return "Model belum siap"

        try:
            context = self._build_context(user_input)
            inputs = self.tokenizer.encode(
                context, return_tensors='pt',
                truncation=True, max_length=self.max_context_tokens,
            ).to(self.device)
            input_length = inputs.shape[1]

            model_max_len = getattr(self.model.config, 'max_position_embeddings', 1024)
            max_new_tokens = min(
                self.gen_config["max_new_tokens"],
                model_max_len - input_length - 10,
            )
            if max_new_tokens <= 0:
                max_new_tokens = 30

            streamer = TextIteratorStreamer(
                self.tokenizer, skip_prompt=True, skip_special_tokens=True
            )

            gen_kwargs = {
                "input_ids": inputs,
                "max_new_tokens": max_new_tokens,
                "temperature": self.gen_config["temperature"],
                "top_p": self.gen_config["top_p"],
                "top_k": self.gen_config["top_k"],
                "repetition_penalty": self.gen_config["repetition_penalty"],
                "no_repeat_ngram_size": self.gen_config["no_repeat_ngram_size"],
                "do_sample": True,
                "pad_token_id": self.tokenizer.eos_token_id,
                "eos_token_id": self.tokenizer.eos_token_id,
                "streamer": streamer,
            }

            thread = Thread(target=self.model.generate, kwargs=gen_kwargs)
            thread.start()

            console.print("[magenta]AI[/]: ", end="")
            chunks: List[str] = []
            for new_text in streamer:
                console.print(rich_escape(new_text), end="")
                chunks.append(new_text)

            console.print()
            console.print()

            response = "".join(chunks)
            response = self._clean_response(response, user_input)
            return response if response else _FALLBACKS["unknown"]

        except torch.cuda.OutOfMemoryError:
            console.print(Theme.error(" GPU OOM. Fallback ke CPU..."))
            try:
                self.model = self.model.to('cpu')
                torch.cuda.empty_cache()
                self.device = 'cpu'
                return self._generate_best_of_n(user_input)
            except Exception as e:
                return f"Error: {str(e)[:100]}"
        except Exception as e:
            error_logger.error(f"Streaming error: {e}")
            return f"Error: {str(e)[:120]}"

    # ==================================================================
    # Model loading (tidak diubah)
    # ==================================================================
    def _load_model(self, model_path: str) -> None:
        is_peft = False
        HAS_PEFT = False
        try:
            from peft import PeftModel
            HAS_PEFT = True
        except ImportError:
            pass

        if os.path.exists(os.path.join(model_path, "adapter_config.json")):
            is_peft = True

        base_info_path = os.path.join(model_path, "base_model_info.json")
        if os.path.exists(base_info_path):
            try:
                with open(base_info_path, 'r') as f:
                    base_info = json.load(f)
                    if base_info.get('use_peft', False):
                        is_peft = True
            except Exception:
                pass

        if is_peft and HAS_PEFT:
            self._load_peft_model(model_path)
        else:
            self._load_full_model(model_path)

    def _load_peft_model(self, model_path: str) -> None:
        try:
            from peft import PeftModel

            base_info_path = os.path.join(model_path, "base_model_info.json")
            if os.path.exists(base_info_path):
                with open(base_info_path, 'r') as f:
                    base_info = json.load(f)
                    base_model_name = base_info.get('base_model', 'gpt2')
            else:
                base_model_name = 'gpt2'

            self.tokenizer = AutoTokenizer.from_pretrained(model_path)
            if self.tokenizer.pad_token is None:
                self.tokenizer.pad_token = self.tokenizer.eos_token

            torch_dtype = torch.float16 if self.device == 'cuda' else torch.float32
            self.model = AutoModelForCausalLM.from_pretrained(
                base_model_name,
                torch_dtype=torch_dtype,
                low_cpu_mem_usage=True,
            )
            self.model = PeftModel.from_pretrained(self.model, model_path)
            self.is_peft_model = True

            self.model = self.model.to(self.device if self.device != 'cpu' else 'cpu')
            console.print(Theme.success(" PEFT model + adapter loaded!"))
        except Exception as e:
            console.print(Theme.warning(f"PEFT load failed: {e}. Fallback ke full model."))
            self._load_full_model(model_path)

    def _load_full_model(self, model_path: str) -> None:
        torch_dtype = torch.float16 if self.device == 'cuda' else torch.float32

        self.tokenizer = AutoTokenizer.from_pretrained(model_path)
        if self.tokenizer.pad_token is None:
            self.tokenizer.pad_token = self.tokenizer.eos_token

        self.model = AutoModelForCausalLM.from_pretrained(
            model_path,
            torch_dtype=torch_dtype,
            low_cpu_mem_usage=True,
        )
        self.model = self.model.to(self.device if self.device != 'cpu' else 'cpu')
        self.is_peft_model = False
        console.print(Theme.success(" Full model loaded!"))

    def _sync_gen_config_from_manager(self) -> None:
        try:
            cfg = self.model_manager.get_generation_config()
            self.max_context_tokens = cfg.get('max_context_length', 768)
            # Hanya temperature/top_p yang di-sync; parameter "aman" tetap
            # pakai default chat.py (max_new_tokens kecil, rep_penalty tinggi).
            if "temperature" in cfg:
                self.gen_config["temperature"] = min(cfg["temperature"], 0.9)
            if "top_p" in cfg:
                self.gen_config["top_p"] = cfg["top_p"]
        except Exception as e:
            debug_logger.debug(f"Sync gen config: {e}")

    # ==================================================================
    # Context building
    # ==================================================================
    def _build_context(self, user_input: str) -> str:
        """Context ketat: system + max 2 turn terakhir + user input."""
        context = f"{self.system_prompt}\n\n"
        total_tokens = len(self.tokenizer.encode(context))

        for h in list(self.conversation_history)[-2:]:
            turn = f"User: {h['user'][:60]}\nAI: {h['ai'][:120]}\n\n"
            turn_tokens = len(self.tokenizer.encode(turn))
            if total_tokens + turn_tokens > self.max_context_tokens - 100:
                break
            context += turn
            total_tokens += turn_tokens

        context += f"User: {user_input}\nAI:"
        return context

    # ==================================================================
    # Stats
    # ==================================================================
    def _update_stats(self, elapsed_ms: float, response: str) -> None:
        self.performance_stats["total_generations"] += 1
        n = self.performance_stats["total_generations"]
        avg = self.performance_stats["avg_time_ms"]
        self.performance_stats["avg_time_ms"] = (avg * (n - 1) + elapsed_ms) / n
        self.performance_stats["max_response_time"] = max(
            self.performance_stats["max_response_time"], elapsed_ms
        )
        self.performance_stats["min_response_time"] = min(
            self.performance_stats["min_response_time"], elapsed_ms
        )
        self.performance_stats["total_tokens_generated"] += len(response.split())

    def _show_stats(self) -> None:
        console.print("\n" + Theme.header(" Chat Performance Statistics:"))

        table = Table(title="Performance Stats", box=box.ROUNDED)
        table.add_column("Metric", style="cyan")
        table.add_column("Value", style="green")

        s = self.performance_stats
        table.add_row("Total Generations", str(s["total_generations"]))
        table.add_row("Avg Response Time", f"{s['avg_time_ms']:.1f}ms")
        table.add_row("Max Response Time", f"{s['max_response_time']:.1f}ms")

        min_t = s["min_response_time"]
        if min_t == float('inf'):
            min_t = 0.0
        table.add_row("Min Response Time", f"{min_t:.1f}ms")

        table.add_row("Total Tokens Generated", f"{s['total_tokens_generated']:,}")
        table.add_row("Total Tokens Processed", f"{s['total_tokens_processed']:,}")
        table.add_row("Fallbacks Triggered", str(s.get("fallbacks_triggered", 0)))
        table.add_row("Regenerations", str(s.get("regenerations", 0)))
        table.add_row("Device", self.device.upper())
        table.add_row("PEFT Model", "Yes" if self.is_peft_model else "No")
        table.add_row("History Length", str(len(self.conversation_history)))
        table.add_row("Streaming", "ON" if self.gen_config["streaming"] else "OFF")

        console.print(table)

        # Config
        table2 = Table(title="Generation Config", box=box.ROUNDED)
        table2.add_column("Parameter", style="cyan")
        table2.add_column("Value", style="green")
        for k, v in self.gen_config.items():
            table2.add_row(k, str(v))
        console.print(table2)

    # ==================================================================
    # History
    # ==================================================================
    def _show_history(self) -> None:
        if not self.conversation_history:
            console.print(Theme.warning("Belum ada riwayat percakapan"))
            return

        console.print("\n" + Theme.header(" Conversation History:"))
        for idx, item in enumerate(self.conversation_history, 1):
            ts = datetime.fromtimestamp(item.get("timestamp", 0)).strftime("%H:%M")
            console.print(f"\n[green][{idx}] User:[/green] [dim]({ts})[/dim]")
            console.print(f"  {rich_escape(str(item['user']))}")
            console.print(f"[magenta]    AI:[/magenta]")
            console.print(f"  {rich_escape(str(item['ai']))}")

    # ==================================================================
    # Session save/load
    # ==================================================================
    def _save_session(self) -> None:
        if not self.conversation_history:
            console.print(Theme.warning("Belum ada percakapan untuk disimpan"))
            return

        ts = datetime.now().strftime("%Y%m%d_%H%M%S")
        out = _SESSION_DIR / f"session_{ts}.json"

        try:
            data = {
                "saved_at": datetime.now().isoformat(),
                "persona": self.system_prompt,
                "gen_config": self.gen_config,
                "stats": self.performance_stats,
                "history": [
                    {
                        "user": h["user"],
                        "ai": h["ai"],
                        "timestamp": h.get("timestamp", 0),
                    }
                    for h in self.conversation_history
                ],
            }
            with open(out, "w", encoding="utf-8") as f:
                json.dump(data, f, indent=2, ensure_ascii=False)
            console.print(Theme.success(f"✓ Session disimpan: {out.name}"))
        except Exception as e:
            console.print(Theme.error(f"Gagal save: {e}"))

    def _list_sessions(self) -> None:
        sessions = sorted(_SESSION_DIR.glob("session_*.json"), reverse=True)
        if not sessions:
            console.print(Theme.warning("Belum ada session tersimpan"))
            return

        table = Table(title="Saved Sessions", box=box.ROUNDED)
        table.add_column("No", style="cyan")
        table.add_column("Nama", style="green")
        table.add_column("Turns", style="blue")
        table.add_column("Waktu", style="dim")

        for i, s in enumerate(sessions, 1):
            try:
                with open(s) as f:
                    data = json.load(f)
                turns = len(data.get("history", []))
                ts = data.get("saved_at", "")[:19].replace("T", " ")
                table.add_row(str(i), s.name, str(turns), ts)
            except Exception:
                table.add_row(str(i), s.name, "?", "?")

        console.print(table)
        console.print(Theme.dim("Load dengan: load <nomor>"))

    def _load_session(self, idx: int) -> None:
        sessions = sorted(_SESSION_DIR.glob("session_*.json"), reverse=True)
        if not sessions:
            console.print(Theme.warning("Belum ada session"))
            return

        if idx < 1 or idx > len(sessions):
            console.print(Theme.error("Nomor tidak valid"))
            return

        path = sessions[idx - 1]
        try:
            with open(path, encoding="utf-8") as f:
                data = json.load(f)

            self.conversation_history.clear()
            for h in data.get("history", []):
                self.conversation_history.append({
                    "user": h["user"],
                    "ai": h["ai"],
                    "timestamp": h.get("timestamp", time.time()),
                })

            if data.get("persona"):
                self.system_prompt = data["persona"]

            console.print(Theme.success(
                f"✓ Session di-load: {len(self.conversation_history)} turns"
            ))
        except Exception as e:
            console.print(Theme.error(f"Gagal load: {e}"))

    # ==================================================================
    # Export chat ke markdown
    # ==================================================================
    def _export_chat(self) -> None:
        if not self.conversation_history:
            console.print(Theme.warning("Belum ada percakapan"))
            return

        ts = datetime.now().strftime("%Y%m%d_%H%M%S")
        out = Path(f"/content/data/chat_export_{ts}.md")

        try:
            lines = [
                "# Chat Export",
                "",
                f"**Tanggal:** {datetime.now().strftime('%Y-%m-%d %H:%M:%S')}",
                f"**Model:** {self.model_manager.get_current_model()}",
                f"**Device:** {self.device.upper()}",
                f"**Total turns:** {len(self.conversation_history)}",
                "",
                "---",
                "",
            ]

            for h in self.conversation_history:
                ts_msg = datetime.fromtimestamp(h.get("timestamp", 0)).strftime("%H:%M:%S")
                lines.append(f"### User ({ts_msg})")
                lines.append("")
                lines.append(h["user"])
                lines.append("")
                lines.append(f"### AI")
                lines.append("")
                lines.append(h["ai"])
                lines.append("")
                lines.append("---")
                lines.append("")

            out.write_text("\n".join(lines), encoding="utf-8")
            console.print(Theme.success(f"✓ Chat di-export: {out}"))
        except Exception as e:
            console.print(Theme.error(f"Gagal export: {e}"))

    # ==================================================================
    # Help
    # ==================================================================
    def _show_help(self) -> None:
        help_text = """
[cyan]KONVERSASI:[/cyan]
  (ketik langsung)   - kirim pesan ke AI

[cyan]RIWAYAT:[/cyan]
  history / hist     - tampilkan riwayat
  undo / u           - hapus turn terakhir
  regen / rg         - jawab ulang pesan terakhir
  clear / xlear      - hapus semua riwayat

[cyan]PARAMETER (runtime):[/cyan]
  temp 0.9           - ubah temperature (0.01-2.0)
  topp 0.85          - ubah top-p (0.0-1.0)
  maxtok 60          - ubah max tokens (10-500)
  attempts 3         - ubah Best-of-N (1-10)
  stream on/off      - aktifkan streaming output

[cyan]PERSONA:[/cyan]
  persona <teks>     - ganti system prompt
  persona reset      - kembali ke default
  persona            - lihat persona sekarang

[cyan]SESSION:[/cyan]
  save / sessions / load N / export

[cyan]LAIN-LAIN:[/cyan]
  stats / s          - statistik performa
  help / hlp         - tampilkan bantuan ini
  exit / q           - keluar

[cyan]TIPS:[/cyan]
- Greeting/thanks/bye dijawab langsung tanpa model (cepat, akurat)
- Best-of-N: model generate 3 jawaban, dipilih yang terbaik
- Kalau jawaban jelek, sistem otomatis fallback ke template sopan
- Gunakan 'regen' kalau ingin jawaban berbeda
- 'attempts 5' bikin lebih variatif (tapi lebih lambat)
"""
        console.print(Panel(help_text.strip(), title="HELP", style="yellow"))