app.py

"""
Gradio RAG -> MCQ app for HuggingFace Spaces
- Upload a PDF
- Chunk + embed using Together embeddings
- Store vectors in Chroma (local) and retrieve
- Call Together chat/completion to generate Vietnamese MCQs in JSON

Drop this file into a new HuggingFace Space (Gradio, Python). Add a requirements.txt (see README below) and set the secret TOGETHER_API_KEY in Space settings.
"""

import os
import json
import uuid
import tempfile
import pdfplumber
from together import Together
import chromadb
from chromadb.config import Settings
import gradio as gr
from typing import List
import shutil

tmp_dir = "./tmp"

# ---------- Config - can be overridden from UI ----------
TOGETHER_API_KEY = os.getenv("TOGETHER_API_KEY")
DEFAULT_EMBEDDING_MODEL = "togethercomputer/m2-bert-80M-8k-retrieval"
DEFAULT_LLM_MODEL = "mistralai/Mixtral-8x7B-Instruct-v0.1"
DEFAULT_CHUNK_SIZE = 1200
DEFAULT_CHUNK_OVERLAP = 200
DEFAULT_K_RETRIEVE = 4
EMBED_BATCH = 64

# instantiate Together client (requires TOGETHER_API_KEY in env / HF Secrets)
if TOGETHER_API_KEY:
    client = Together(api_key=TOGETHER_API_KEY)
else:
    # allow local testing if user wants to set env var later
    client = None

# -------- PDF -> text ----------
def extract_text_from_pdf(path: str) -> str:
    text_parts = []
    with pdfplumber.open(path) as pdf:
        for page in pdf.pages:
            page_text = page.extract_text()
            if page_text:
                text_parts.append(page_text)
    return "\n\n".join(text_parts)

# -------- simple chunker ----------
def chunk_text(text: str, chunk_size=DEFAULT_CHUNK_SIZE, overlap=DEFAULT_CHUNK_OVERLAP) -> List[str]:
    chunks = []
    start = 0
    L = len(text)
    while start < L:
        end = min(L, start + chunk_size)
        chunk = text[start:end].strip()
        chunks.append(chunk)
        start = end - overlap
        if start < 0:
            start = 0
        if start >= L:
            break
    return chunks

# -------- embeddings (Together) with batching ----------
def embed_texts(texts: List[str], model=DEFAULT_EMBEDDING_MODEL):
    if client is None:
        raise RuntimeError("Together client not initialized. Set TOGETHER_API_KEY in environment or Space secrets.")
    embeddings = []
    for i in range(0, len(texts), EMBED_BATCH):
        batch = texts[i:i+EMBED_BATCH]
        resp = client.embeddings.create(input=batch, model=model)
        # resp.data is list; each item has .embedding
        for item in resp.data:
            embeddings.append(item.embedding)
    return embeddings

# -------- chroma vectorstore setup / helpers ----------
def build_chroma_collection(name="pdf_docs", persist_directory="./chroma_db"):
    # On Spaces, writes to the repo may be limited; chroma will attempt to use the path provided.
    client_chroma = chromadb.Client(Settings(chroma_db_impl="duckdb+parquet", persist_directory=persist_directory))
    # create or get collection
    try:
        collection = client_chroma.get_collection(name)
    except Exception:
        collection = client_chroma.create_collection(name)
    return client_chroma, collection

def add_documents_to_vectorstore(collection, chunks: List[str], embeddings: List[List[float]]):
    ids = [f"doc_{i}" for i in range(len(chunks))]
    metadatas = [{"chunk_index": i} for i in range(len(chunks))]
    # If collection already has docs with same ids, Chroma will append; it's common to recreate collection per-upload.
    collection.add(ids=ids, documents=chunks, metadatas=metadatas, embeddings=embeddings)

# -------- retrieve top-k using chroma ----------
def retrieve_relevant_chunks(collection, query: str, k=DEFAULT_K_RETRIEVE, embedding_model=DEFAULT_EMBEDDING_MODEL):
    q_emb = embed_texts([query], model=embedding_model)[0]
    result = collection.query(query_embeddings=[q_emb], n_results=k, include=["documents", "metadatas", "distances"])
    docs = result["documents"][0]
    metas = result["metadatas"][0]
    distances = result["distances"][0]
    return list(zip(docs, metas, distances))

# -------- prompt template (Vietnamese) ----------
MCQ_PROMPT_VI = """
Bạn là một chuyên gia soạn câu hỏi trắc nghiệm (MCQ). SỬ DỤNG CHỈ các đoạn ngữ cảnh được cung cấp dưới đây (KHÔNG suy diễn/không thêm thông tin ngoài ngữ cảnh).
Tạo **một** câu hỏi trắc nghiệm có 4 lựa chọn (A, B, C, D), chỉ ra đáp án đúng (A/B/C/D) và viết 1 câu giải thích ngắn (1-2 câu).
**Bắt buộc:** output PHẢI LÀ **JSON duy nhất** theo schema sau (không có văn bản nào khác ngoài JSON):

{{
  "question_id": "<mã duy nhất>",
  "question": "<câu hỏi bằng tiếng Việt>",
  "options": [
    {{ "label": "A", "text": "..." }},
    {{ "label": "B", "text": "..." }},
    {{ "label": "C", "text": "..." }},
    {{ "label": "D", "text": "..." }}
  ],
  "answer": "A",
  "explanation": "<giải thích ngắn bằng tiếng Việt>",
  "source_chunks": [ "<chunk_index hoặc đoạn trích ngắn>", ... ]
}}

Ví dụ đầu ra (một mẫu JSON đúng; chỉ để mô tả định dạng):
{{
  "question_id": "q_0001",
  "question": "Nguyên tố nào là thành phần chính của vỏ trái đất?",
  "options": [
    {{ "label": "A", "text": "Sắt" }},
    {{ "label": "B", "text": "Oxi" }},
    {{ "label": "C", "text": "Cacbon" }},
    {{ "label": "D", "text": "Nitơ" }}
  ],
  "answer": "B",
  "explanation": "Oxi là nguyên tố phong phú nhất trong vỏ trái đất, chủ yếu trong các oxit và khoáng vật.",
  "source_chunks": [ "chunk_3" ]
}}

Đây là các đoạn ngữ cảnh (chỉ được phép dùng những đoạn này để soạn câu hỏi):
{context}

Hãy viết câu hỏi rõ ràng, không gây mơ hồ. Đảm bảo distractor (đáp án sai) là hợp lý và gây nhầm lẫn cho người học.
"""

# -------- call Together chat/completion ----------
def generate_mcq_with_rag(question_seed: str, retrieved_chunks, llm_model=DEFAULT_LLM_MODEL, temperature=0.0):
    if client is None:
        raise RuntimeError("Together client not initialized. Set TOGETHER_API_KEY in environment or Space secrets.")

    context = ""
    for i, (doc_text, meta, dist) in enumerate(retrieved_chunks):
        snippet = doc_text.replace("\n", " ").strip()
        context += f"[chunk_{meta.get('chunk_index', i)}] {snippet}\n\n"

    prompt = MCQ_PROMPT_VI.format(context=context)
    full_user = f"Yêu cầu (chủ đề / seed): {question_seed}\n\n{prompt}"

    messages = [
        {"role": "system", "content": "Bạn là một chuyên gia soạn câu hỏi trắc nghiệm bằng tiếng Việt. Chỉ trả về JSON, KHÔNG có lời giải thích thêm."},
        {"role": "user", "content": full_user},
    ]

    resp = client.chat.completions.create(
        model=llm_model,
        messages=messages,
        temperature=temperature,
    )
    out = resp.choices[0].message.content

    # try to parse JSON, fallback to extracting first {...}
    try:
        parsed = json.loads(out)
    except Exception:
        start = out.find("{")
        end = out.rfind("}")
        if start != -1 and end != -1:
            try:
                parsed = json.loads(out[start:end+1])
            except Exception:
                parsed = None
        else:
            parsed = None

    # ensure question_id exists
    if parsed and isinstance(parsed, dict):
        if not parsed.get("question_id"):
            parsed["question_id"] = f"q_{uuid.uuid4().hex[:8]}"

    return parsed, out

# -------- high-level runner used by Gradio ----------
def generate_mcqs_from_pdf(pdf_path: str, seeds: List[str], questions_per_seed=1, chunk_size=DEFAULT_CHUNK_SIZE,
                            chunk_overlap=DEFAULT_CHUNK_OVERLAP, k_retrieve=DEFAULT_K_RETRIEVE,
                            embedding_model=DEFAULT_EMBEDDING_MODEL, llm_model=DEFAULT_LLM_MODEL,
                            temperature=0.0, persist_directory="./chroma_db"):
    text = extract_text_from_pdf(pdf_path)
    chunks = chunk_text(text, chunk_size=chunk_size, overlap=chunk_overlap)

    # embed
    chunk_embeddings = embed_texts(chunks, model=embedding_model)

    # build vectorstore (recreate to avoid old data)
    chroma_client, collection = build_chroma_collection(name="pdf_docs", persist_directory=persist_directory)
    try:
        collection.delete()
        collection = chroma_client.create_collection("pdf_docs")
    except Exception:
        # some backends will raise; ignore and continue
        pass
    add_documents_to_vectorstore(collection, chunks, chunk_embeddings)

    results = []
    for seed in seeds:
        for i in range(questions_per_seed):
            retrieved = retrieve_relevant_chunks(collection, seed, k=k_retrieve, embedding_model=embedding_model)
            parsed, raw = generate_mcq_with_rag(seed, retrieved, llm_model=llm_model, temperature=temperature)
            if parsed is None:
                item = {"seed": seed, "ok": False, "raw": raw}
            else:
                item = {"seed": seed, "ok": True, "mcq": parsed}
            results.append(item)

    return results

# -------- Gradio UI ----------

def save_uploaded_file(uploaded) -> str:
    """
    uploaded may be:
      - Path string (when running locally in some setups)
      - File-like object with .name
      - tuple/list returned by gradio in some versions
    Returns saved file path.
    """
    if uploaded is None:
        raise ValueError("No file uploaded.")
    # normalize to path
    if isinstance(uploaded, str) and os.path.exists(uploaded):
        src = uploaded
    elif hasattr(uploaded, "name") and os.path.exists(uploaded.name):
        src = uploaded.name
    elif isinstance(uploaded, (tuple, list)) and len(uploaded) > 0:
        # sometimes gradio returns (tempfile_path, original_name)
        cand = uploaded[0]
        if isinstance(cand, str) and os.path.exists(cand):
            src = cand
        else:
            # fallback: try bytes
            src = None
    else:
        src = None

    dest_path = os.path.join(tmp_dir, os.path.basename(src) if src else "uploaded_doc")
    if src:
        shutil.copy(src, dest_path)
        return dest_path

    # last-resort: if 'uploaded' is bytes-like
    try:
        data = uploaded.read()
    except Exception:
        # try treat as bytes
        data = uploaded if isinstance(uploaded, (bytes, bytearray)) else None

    if data is None:
        raise ValueError("Could not handle uploaded file type.")
    with open(dest_path, "wb") as f:
        f.write(data)
    return dest_path

def ui_run(pdf_file, seeds_text, questions_per_seed, k_retrieve, chunk_size, chunk_overlap,
           embedding_model, llm_model, temperature):
    if pdf_file is None:
        return "", None

    # save uploaded file to temp path
    try:
        # Clear tmp folder and recreate
        if os.path.exists(tmp_dir):
            shutil.rmtree(tmp_dir)
        os.makedirs(tmp_dir)
    except Exception:
        pass

    print(f"created {tmp_dir}")

    try:
        local_path = save_uploaded_file(pdf_file)
    except Exception as e:
        return {"error": f"Failed saving uploaded file: {e}"}

    print(f"uploaded file {local_path}")

    seeds = [s.strip() for s in seeds_text.split(",") if s.strip()]
    if not seeds:
        seeds = ["Lấy câu hỏi tổng quát về tài liệu"]

    print(f"seeds: {seeds}")

    print("generating mcqs")
    try:
        results = generate_mcqs_from_pdf(
            pdf_path=local_path,
            seeds=seeds,
            questions_per_seed=questions_per_seed,
            chunk_size=chunk_size,
            chunk_overlap=chunk_overlap,
            k_retrieve=k_retrieve,
            embedding_model=embedding_model,
            llm_model=llm_model,
            temperature=temperature,
            persist_directory="./chroma_db"
        )
    except Exception as e:
        return f"Lỗi khi sinh MCQ: {e}", None
    print("mcqs generated")

    out_json = json.dumps(results, ensure_ascii=False, indent=2)
    # write output file for download
    out_file = os.path.join(tmp_dir, "mcq_output.json")
    with open(out_file, "w", encoding="utf-8") as f:
        f.write(out_json)
    print("json output dumped")

    return out_json, out_file

with gr.Blocks(title="RAG -> MCQ (Tiếng Việt)") as demo:
    gr.Markdown("# RAG -> MCQ Generator (Tiếng Việt)\nUpload PDF, set seeds (phân tách bằng dấu phẩy), và nhấn Generate.\nOutputs: JSON trả về các câu hỏi trắc nghiệm.)")

    with gr.Row():
        with gr.Column(scale=1):
            pdf_in = gr.File(label="Upload PDF")
            seeds_in = gr.Textbox(label="Seeds (chủ đề), phân tách bằng dấu phẩy", value="lập trình hướng đối tượng, kế thừa")
            questions_per_seed = gr.Slider(label="Questions per seed", minimum=1, maximum=5, step=1, value=1)
            k_retrieve = gr.Slider(label="K retrieve (số đoạn liên quan)", minimum=1, maximum=10, step=1, value=DEFAULT_K_RETRIEVE)
            chunk_size = gr.Number(label="Chunk size (chars)", value=DEFAULT_CHUNK_SIZE)
            chunk_overlap = gr.Number(label="Chunk overlap (chars)", value=DEFAULT_CHUNK_OVERLAP)
            embedding_model = gr.Textbox(label="Embedding model", value=DEFAULT_EMBEDDING_MODEL)
            llm_model = gr.Textbox(label="LLM model", value=DEFAULT_LLM_MODEL)
            temperature = gr.Slider(label="Temperature", minimum=0.0, maximum=1.0, step=0.05, value=0.0)
            btn = gr.Button("Generate MCQs")
        with gr.Column(scale=1):
            out_text = gr.Textbox(label="Raw JSON output", lines=20)
            out_file = gr.File(label="Download JSON")

    btn.click(fn=ui_run, inputs=[pdf_in, seeds_in, questions_per_seed, k_retrieve, chunk_size, chunk_overlap,
                                 embedding_model, llm_model, temperature], outputs=[out_text, out_file])

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