model_utils.py

# model_utils.py
from typing import List, Optional
import re

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
from sentence_transformers import SentenceTransformer
from sentence_transformers.util import cos_sim

import qa_store
from loader import load_curriculum, load_manual_qa, rebuild_combined_qa

# -----------------------------
# Base chat model
# -----------------------------
MODEL_NAME = "SeaLLMs/SeaLLMs-v3-1.5B-Chat"
MAX_CONTEXT_ENTRIES = 3  # how many textbook chunks to retrieve per question

tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)

device = "cuda" if torch.cuda.is_available() else "cpu"

model = AutoModelForCausalLM.from_pretrained(
    MODEL_NAME,
    torch_dtype=torch.float32,
).to(device)
model.eval()

# -----------------------------
# Embedding model for retrieval
# -----------------------------
EMBED_MODEL_NAME = "sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2"
embed_model = SentenceTransformer(EMBED_MODEL_NAME)
# (optional) move embedding model to same device; OK to leave on CPU if you want
embed_model = embed_model.to(device)


# NOTE: called once after load_curriculum() to precompute embeddings.
# If you ever reload curriculum at runtime, call _build_entry_embeddings() again.
def _build_entry_embeddings() -> None:
    """
    Build embeddings for each textbook entry using title + summary + text
    and store them in qa_store.TEXT_EMBEDDINGS.
    """
    if not qa_store.ENTRIES:
        qa_store.TEXT_EMBEDDINGS = None
        return

    texts = []
    for e in qa_store.ENTRIES:
        title = e.get("title", "") or ""
        summary = e.get("summary", "") or ""
        text = e.get("text", "") or ""
        combined = f"{title}\n{summary}\n{text}"
        texts.append(combined)

    qa_store.TEXT_EMBEDDINGS = embed_model.encode(
        texts,
        convert_to_tensor=True,
        show_progress_bar=False,
    )


# -----------------------------
# Load data once at import time
# -----------------------------
load_curriculum()
load_manual_qa()
rebuild_combined_qa()
_build_entry_embeddings()

SYSTEM_PROMPT = (
    "ທ່ານແມ່ນຜູ້ຊ່ວຍເຫຼືອດ້ານປະຫວັດສາດຂອງປະເທດລາວ "
    "ສໍາລັບນັກຮຽນຊັ້ນ ມ.1. "
    "ຕອບແຕ່ພາສາລາວ ໃຫ້ຕອບສັ້ນໆ 2–3 ປະໂຫຍກ ແລະເຂົ້າໃຈງ່າຍ. "
    "ໃຫ້ອີງຈາກຂໍ້ມູນຂ້າງລຸ່ມນີ້ເທົ່ານັ້ນ. "
    "ຖ້າຂໍ້ມູນບໍ່ພຽງພໍ ຫຼືບໍ່ຊັດເຈນ ໃຫ້ບອກວ່າບໍ່ແນ່ໃຈ."
)


def retrieve_context(question: str, max_entries: int = MAX_CONTEXT_ENTRIES) -> str:
    """
    Embedding-based retrieval over textbook entries.
    Falls back to first entries if embeddings are missing.
    """
    if not qa_store.ENTRIES:
        return qa_store.RAW_KNOWLEDGE

    if qa_store.TEXT_EMBEDDINGS is None:
        top_entries = qa_store.ENTRIES[:max_entries]
    else:
        # 1) Encode the question
        q_vec = embed_model.encode(
            question,
            convert_to_tensor=True,
            show_progress_bar=False,
        )

        # 2) Cosine similarity with all entry embeddings
        sims = cos_sim(q_vec, qa_store.TEXT_EMBEDDINGS)[0]  # shape [N]

        # 3) Pick top-k indices
        k = min(max_entries, len(qa_store.ENTRIES))
        _, top_indices = torch.topk(sims, k=k)

        # 4) Map indices back to entries
        top_entries = [qa_store.ENTRIES[i] for i in top_indices.tolist()]

    # Build context string for the prompt
    context_blocks: List[str] = []
    for e in top_entries:
        header = (
            f"[ຊັ້ນ {e.get('grade','')}, "
            f"ບົດ {e.get('chapter','')}, "
            f"ຫົວຂໍ້ {e.get('section','')} – {e.get('title','')}]"
        )
        context_blocks.append(f"{header}\n{e.get('text','')}")

    return "\n\n".join(context_blocks)


def _format_history(history: Optional[List]) -> str:
    """
    Convert last few chat turns into a Lao conversation snippet
    to give the model context for follow-up questions.
    Gradio history format: [[user_msg, bot_msg], [user_msg, bot_msg], ...]
    """
    if not history:
        return ""

    # keep only the last 3 turns to avoid very long prompts
    recent = history[-3:]

    lines: List[str] = []
    for turn in recent:
        if not isinstance(turn, (list, tuple)) or len(turn) != 2:
            continue
        user_msg, bot_msg = turn
        lines.append(f"ນັກຮຽນ: {user_msg}")
        lines.append(f"ອາຈານ AI: {bot_msg}")

    if not lines:
        return ""

    joined = "\n".join(lines)
    return f"ປະຫວັດການສົນທະນາກ່ອນໜ້າ:\n{joined}\n\n"


def build_prompt(question: str, history: Optional[List] = None) -> str:
    context = retrieve_context(question, max_entries=MAX_CONTEXT_ENTRIES)
    history_block = _format_history(history)

    return f"""{SYSTEM_PROMPT}

{history_block}ຂໍ້ມູນອ້າງອີງ:
{context}

ຄຳຖາມ: {question}

ຄຳຕອບດ້ວຍພາສາລາວ:"""


def generate_answer(question: str, history: Optional[List] = None) -> str:
    prompt = build_prompt(question, history)
    inputs = tokenizer(prompt, return_tensors="pt").to(device)

    with torch.no_grad():
        outputs = model.generate(
            **inputs,
            max_new_tokens=160,
            do_sample=False,
        )

    generated_ids = outputs[0][inputs["input_ids"].shape[1]:]
    answer = tokenizer.decode(generated_ids, skip_special_tokens=True).strip()

    # Enforce 2–3 sentence answers for M.1 students
    sentences = re.split(r"(?<=[\.?!…])\s+", answer)
    short_answer = " ".join(sentences[:3]).strip()
    return short_answer if short_answer else answer


def answer_from_qa(question: str) -> Optional[str]:
    """
    1) Exact match in QA_INDEX
    2) Fuzzy match via word overlap with ALL_QA_KNOWLEDGE
    """
    norm_q = qa_store.normalize_question(question)
    if not norm_q:
        return None

    # Exact match
    if norm_q in qa_store.QA_INDEX:
        return qa_store.QA_INDEX[norm_q]

    # Fuzzy match
    q_terms = [t for t in norm_q.split(" ") if len(t) > 1]
    if not q_terms:
        return None

    best_score = 0
    best_answer: Optional[str] = None

    for item in qa_store.ALL_QA_KNOWLEDGE:
        stored_terms = [t for t in item["norm_q"].split(" ") if len(t) > 1]
        overlap = sum(1 for t in q_terms if t in stored_terms)
        if overlap > best_score:
            best_score = overlap
            best_answer = item["a"]

    # require at least 2 overlapping words to accept fuzzy match
    if best_score >= 2 and best_answer is not None:
        # optional: log when fuzzy match is used
        print(f"[FUZZY MATCH] score={best_score} -> {best_answer[:50]!r}")
        return best_answer

    return None


def laos_history_bot(message: str, history: List) -> str:
    """
    Main chatbot function for Student tab (Gradio ChatInterface).
    """
    if not message.strip():
        return "ກະລຸນາພິມຄໍາຖາມກ່ອນ."

    # 1) Try exact / fuzzy Q&A first
    direct = answer_from_qa(message)
    if direct:
        return direct

    # 2) Fall back to LLM + retrieved context
    try:
        answer = generate_answer(message, history)
    except Exception as e:  # noqa: BLE001
        return f"ລະບົບມີບັນຫາ: {e}"

    return answer