app.py

import os
import csv
import uuid
from datetime import datetime
from typing import List, Tuple
import torch
import gradio as gr
from filelock import FileLock
from huggingface_hub import HfApi
from transformers import (
    AutoTokenizer,
    AutoModelForCausalLM,
    pipeline,
    StoppingCriteria,
    StoppingCriteriaList,
)
from peft import PeftModel
import tempfile
import pandas as pd
from datasets import load_dataset

# =========================
# ⚙️ Config
# =========================
MAX_HISTORY_TURNS = 10
MAX_PROMPT_TOKENS = 1024
MAX_NEW_TOKENS = 60

LOG_DIR = "logs"
os.makedirs(LOG_DIR, exist_ok=True)
LOCK_PATH = os.path.join(LOG_DIR, ".lock")

HF_TOKEN = os.environ.get("HF_TOKEN")
SPACE_ID = os.environ.get("SPACE_ID")

MODEL_ID = "hparten/prob1_qlora_math_student"

# =========================
# 🔠 Model + Tokenizer
# =========================
model = AutoModelForCausalLM.from_pretrained(
    MODEL_ID,
    device_map="auto",
)

tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
tokenizer.pad_token = tokenizer.eos_token

pipe = pipeline(
    "text-generation",
    model=model,
    tokenizer=tokenizer,
    dtype=torch.float16,
    device_map="auto",
)

# =========================
# 🧩 Strategy Explanations
# =========================
strategy_explanations = {
    "friendly": "You add on from 41 until you get to 84, usually by counting by 10s, 20s, or 40, then ones.",
    "differencing": "You difference the ones or tens separately during any part of your answer.",
    "subtraction": "You turn the problem into a subtraction: 84 minus 41 equals blank to find the missing addend.",
}

# =========================
# 🧠 Build System Prompt
# =========================
def build_system_block(problem_prefix, strategy):
    problem_text = "41 plus blank equals 84"
    strat_key = strategy.lower()
    strat_expl = strategy_explanations.get(strat_key, "Use the named strategy to explain your steps clearly.")
    strategy_tag = f"<strategy_{strat_key}>"
    problem_tag = f"<{problem_prefix.lower()}>"

    system_text = (
        f"<system>\n"
        f"You are the student in a math dialogue.\n"
        f"Solving the PROBLEM: {problem_tag} - {problem_text}\n"
        f"Using the STRATEGY: {strategy_tag} — {strat_expl}\n"
        f"Return EXACTLY one sentence inside <student> ... </student>."
        f"Do NOT ask questions or include teacher text.\n"
        f"Mention the strategy implicity only if natural.\n"
        f"</system>\n"
    )
    return system_text.strip()

# =========================
# 🧾 Logging
# =========================
#CSV_HEADERS = ["timestamp", "session_id", "username", "strategy", "teacher", "student"]
#
#def _append_csv(path, row):
#    with FileLock(LOCK_PATH):
#        file_exists = os.path.exists(path)
#        with open(path, "a", newline="", encoding="utf-8") as f:
#            w = csv.writer(f)
#            if not file_exists:
#                w.writerow(CSV_HEADERS)
#            w.writerow(row)
#
#def log_turn(session_id, username, strategy, teacher_msg, student_msg):
#    row = [datetime.now().isoformat(timespec="seconds"), session_id, username, strategy, #teacher_msg, student_msg]
#    per_session = os.path.join(LOG_DIR, f"chat_{session_id}.csv")
#    _append_csv(per_session, row)

# =========================
# 🧩 Prompt builder
# =========================
def build_prompt(strategy, history, teacher_question, tokenizer, problem_prefix="Problem_1"):
    base_system_prompt = build_system_block(problem_prefix, strategy)

    turns = []
    for tq, sa in history[-MAX_HISTORY_TURNS:]:
        turns.append(f"<teacher> {tq} </teacher> <student> {sa} </student>")

    full_prompt = base_system_prompt + "\n" + " ".join(turns)
    full_prompt += f"<teacher> {teacher_question} </teacher>\n<student>"

    while len(tokenizer.encode(full_prompt, add_special_tokens=False)) > MAX_PROMPT_TOKENS and len(turns) > 0:
        turns.pop(0)
        convo_block = " ".join(turns)
        full_prompt = base_system_prompt + convo_block + f"<teacher> {teacher_question} </teacher>"

    return full_prompt.strip()

# =========================================================
# ❌ Banned Tokens (prevents teacher drift)
# =========================================================
def make_bad_words_ids(tokenizer, words: List[str]) -> List[List[int]]:
    """Safely constructs bad_words_ids for special and normal tokens."""
    out = []
    for w in words:
        if w in tokenizer.all_special_tokens:
            tid = tokenizer.convert_tokens_to_ids(w)
            if tid != tokenizer.unk_token_id:
                out.append([tid])
        else:
            toks = tokenizer.encode(w, add_special_tokens=False)
            if toks:
                out.append(toks)
    return out

bad_words_ids = make_bad_words_ids(
    tokenizer,
    ["<teacher>", "</teacher>", "<system>", "</system>", "Teacher:", "teacher:"]
)
eos_id = tokenizer.convert_tokens_to_ids("</student>")

# =========================
# ☁️ In-Memory Logging + HF Upload
# =========================

HF_DATASET_REPO = "hparten/math_chatbot_logs"

api = HfApi()
session_logs = {}       # session_id -> list of turns
last_activity = {}      # session_id -> timestamp

def add_turn_to_memory(session_id, username, strategy, teacher_msg, student_msg):
    """Store one turn in memory."""
    from datetime import datetime
    row = {
        "timestamp": datetime.now().isoformat(timespec="seconds"),
        "session_id": session_id,
        "username": username,
        "strategy": strategy,
        "teacher": teacher_msg,
        "student": student_msg,
    }
    session_logs.setdefault(session_id, []).append(row)
    update_activity(session_id)

def update_activity(session_id):
    import time
    last_activity[session_id] = time.time()

def flush_session_to_hub(session_id):
    """Upload session logs to Hugging Face dataset as a single Parquet file."""
    if session_id not in session_logs or not session_logs[session_id]:
        print(f"[flush] No logs found for session {session_id}")
        return

    df = pd.DataFrame(session_logs[session_id])
    del session_logs[session_id]

    try:
        ds = load_dataset(HF_DATASET_REPO, split="train", token=HF_TOKEN)
        existing = ds.to_pandas()
        combined = pd.concat([existing, df], ignore_index=True)
    except Exception:
        combined = df

    with tempfile.NamedTemporaryFile("wb", delete=False, suffix=".parquet") as tmp:
        combined.to_parquet(tmp.name, index=False)
        tmp_path = tmp.name

    api.upload_file(
        path_or_fileobj=tmp_path,
        path_in_repo="chat_logs.parquet",
        repo_id=HF_DATASET_REPO,
        repo_type="dataset",
        token=HF_TOKEN,
    )
    os.remove(tmp_path)
    print(f"[flush] Uploaded session {session_id} to HF dataset.")

# =========================
# 🤖 Generation
# =========================
def generate_response(teacher_question, username, history, session_id, strategy):
    prompt = build_prompt(strategy, history, teacher_question, tokenizer)

    out = pipe(
        prompt,
        max_new_tokens=MAX_NEW_TOKENS,
        do_sample=True,
        temperature=0.4,
        top_p=0.9,
        repetition_penalty=1.05,
        no_repeat_ngram_size=6,
        pad_token_id=tokenizer.pad_token_id or tokenizer.eos_token_id,
        eos_token_id=eos_id,
        bad_words_ids=bad_words_ids,
        return_full_text=False,
    )

    out_text = out[0]["generated_text"]
    if "<student>" in out_text and "</student>" in out_text:
        student_reply = out_text.split("<student>", 1)[1].split("</student>", 1)[0].strip()
    else:
        student_reply = out_text.strip()

    # Force single-sentence cleanup
    student_reply = student_reply.split(".")[0].strip() + "."

    history.append((teacher_question, student_reply))
    add_turn_to_memory(session_id, username, strategy, teacher_question, student_reply)

    return student_reply, history

# =========================
# ☁️ Flush session logs to Hugging Face Hub
# =========================
#
#def flush_session_to_hub(session_id):
#    """Append this session to one Parquet file in the private HF dataset."""
#    if session_id not in session_logs or not session_logs[session_id]:
#        print(f"[flush] No logs found for session {session_id}")
#        return
#
#    df = pd.DataFrame(session_logs[session_id])
#    del session_logs[session_id]
#
#    try:
#        ds = load_dataset(HF_DATASET_REPO, split="train", token=HF_TOKEN)
#        existing = ds.to_pandas()
#        combined = pd.concat([existing, df], ignore_index=True)
#    except Exception:
#        combined = df
#
#    with tempfile.NamedTemporaryFile("wb", delete=False, suffix=".parquet") as tmp:
#        combined.to_parquet(tmp.name, index=False)
#        tmp_path = tmp.name
#
#    api.upload_file(
#        path_or_fileobj=tmp_path,
#        path_in_repo="chat_logs.parquet",
#        repo_id=HF_DATASET_REPO,
#        repo_type="dataset",
#        token=HF_TOKEN,
#    )
#
#    os.remove(tmp_path)
#    print(f"[flush] Uploaded session {session_id} to HF dataset.")

# =========================
# Inactivity flush
# =========================

import threading, time

INACTIVITY_LIMIT = 600  # 10 minutes

def check_inactivity_loop():
    while True:
        now = time.time()
        inactive = [sid for sid, ts in last_activity.items() if now - ts > INACTIVITY_LIMIT]
        for sid in inactive:
            try:
                flush_session_to_hub(sid)
                del last_activity[sid]
            except Exception as e:
                print(f"[auto-flush-error] {sid}: {e}")
        time.sleep(60)

threading.Thread(target=check_inactivity_loop, daemon=True).start()

# =========================
# 🖥 Gradio UI
# =========================
def on_send(teacher_question, username, strategy_choice, history, session_id):
    if not session_id:
        session_id = uuid.uuid4().hex[:12]
    if history is None:
        history = []
    if not username.strip():
        gr.Warning("Please enter your name before starting the chat.")
        return history, history, "", session_id
    if not teacher_question.strip():
        gr.Warning("Please type a question for the student before sending.")
        return history, history, "", session_id

    student_reply, history = generate_response(
        teacher_question.strip(),
        username.strip(),
        history,
        session_id,
        strategy_choice.lower()
    )

    msgs = []
    for t, s in history[-MAX_HISTORY_TURNS:]:
        msgs.append({"role": "user", "content": t})
        msgs.append({"role": "assistant", "content": s})

    return msgs, history, "", session_id

def on_reset(chat, history, teacher_q, session_id):
    """Flush the current session before resetting."""
    if session_id:
        try:
            flush_session_to_hub(session_id)
            print(f"[manual flush] Uploaded session {session_id} to HF dataset.")
        except Exception as e:
            print(f"[manual flush error] {session_id}: {e}")

    return [], [], "", uuid.uuid4().hex[:12]

# =========================
# 🚀 Gradio App
# =========================
with gr.Blocks(title="Elementary Math Student Chatbot") as demo:
    gr.Markdown("## 🧮 Practice Eliciting Student Thinking (Prototype)")
    gr.Markdown(
        "You are an elementary math teacher exploring a student's reasoning for **41 + ___ = 84**."
        "\nAsk questions and see how the student explains their thinking."
    )

    with gr.Row():
        username = gr.Textbox(label="👤 Your Name (first last)", placeholder="Enter your name...")
        strategy_choice = gr.Dropdown(
            ["friendly", "differencing", "subtraction"],
            value="Choose one",
            label="🧩 Student Strategy"
        )
        reset_btn = gr.Button("🔄 Start Over", variant="secondary")

    teacher_q = gr.Textbox(label="👩‍🏫 Teacher Question", placeholder="Ask the student a question…")
    chat = gr.Chatbot(label="💬 Chat", type="messages")
    state_history = gr.State([])
    state_session = gr.State("")

    send = gr.Button("Send", variant="primary")

    send.click(
        on_send,
        inputs=[teacher_q, username, strategy_choice, state_history, state_session],
        outputs=[chat, state_history, teacher_q, state_session],
    )

    reset_btn.click(
    on_reset,
    inputs=[chat, state_history, teacher_q, state_session],
    outputs=[chat, state_history, teacher_q, state_session],
)


if __name__ == "__main__":
    demo.queue()
    demo.launch(server_name="0.0.0.0", server_port=7860, ssr_mode=False)