app.py

import os, re, uuid
from pathlib import Path

_CACHE_DIR = "/data/hf_cache"
os.makedirs(_CACHE_DIR, exist_ok=True)
os.environ["HF_HOME"]             = _CACHE_DIR
os.environ["TRANSFORMERS_CACHE"]  = _CACHE_DIR
os.environ["HF_DATASETS_CACHE"]   = _CACHE_DIR

import gradio as gr
import chromadb
from chromadb.config import Settings
import fitz
from docx import Document as DocxDocument
from sentence_transformers import SentenceTransformer
from transformers import AutoTokenizer, AutoModelForCausalLM, pipeline
import torch

EMBED_MODEL_ID  = "intfloat/multilingual-e5-small"
LLM_MODEL_ID    = "Qwen/Qwen2.5-0.5B-Instruct"
COLLECTION_NAME = "rag_docs"
CHROMA_PATH     = "/data/chromadb"
CHUNK_SIZE, CHUNK_OVERLAP, TOP_K, MAX_NEW_TOKENS = 512, 64, 4, 512
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"

print(f"[INFO] Loading embedding: {EMBED_MODEL_ID}")
embed_model = SentenceTransformer(EMBED_MODEL_ID, device=DEVICE)

print(f"[INFO] Loading LLM: {LLM_MODEL_ID}")
tokenizer = AutoTokenizer.from_pretrained(LLM_MODEL_ID)
llm_model = AutoModelForCausalLM.from_pretrained(
    LLM_MODEL_ID,
    dtype=torch.float16 if DEVICE == "cuda" else torch.float32,
    device_map="auto" if DEVICE == "cuda" else None,
)
llm_pipeline = pipeline(
    "text-generation", model=llm_model, tokenizer=tokenizer,
    device=DEVICE if DEVICE == "cpu" else None,
)

# ── ChromaDB ──────────────────────────────────────────────────────────────────
def _make_chroma():
    if os.path.exists("/data"):
        os.makedirs(CHROMA_PATH, exist_ok=True)
        return chromadb.PersistentClient(path=CHROMA_PATH)
    return chromadb.Client(Settings(anonymized_telemetry=False))

chroma_client = _make_chroma()
collection = chroma_client.get_or_create_collection(
    name=COLLECTION_NAME, metadata={"hnsw:space": "cosine"})

# ── Document loaders ──────────────────────────────────────────────────────────
def load_pdf(p):
    doc = fitz.open(p)
    t = "\n\n".join(pg.get_text() for pg in doc)
    doc.close()
    return t

def load_docx(p):
    return "\n\n".join(para.text for para in DocxDocument(p).paragraphs if para.text.strip())

def load_text(p):
    return open(p, encoding="utf-8", errors="ignore").read()

def load_document(p):
    ext = Path(p).suffix.lower()
    if ext == ".pdf":            return load_pdf(p)
    if ext == ".docx":           return load_docx(p)
    if ext in (".txt", ".md"):   return load_text(p)
    raise ValueError(f"Unsupported: {ext}")

# ── Chunking ──────────────────────────────────────────────────────────────────
def split_text(text):
    text = re.sub(r"\n{3,}", "\n\n", text).strip()
    chunks, start = [], 0
    while start < len(text):
        end = start + CHUNK_SIZE
        if end >= len(text):
            chunks.append(text[start:])
            break
        sp = end
        for sep in ["\n\n", "\n", ". "]:
            idx = text.rfind(sep, start + CHUNK_OVERLAP, end)
            if idx != -1:
                sp = idx + len(sep)
                break
        chunks.append(text[start:sp])
        start = sp - CHUNK_OVERLAP
    return [c.strip() for c in chunks if len(c.strip()) > 30]

# ── Embeddings ────────────────────────────────────────────────────────────────
def embed_passages(texts):
    return embed_model.encode(
        [f"passage: {t}" for t in texts], show_progress_bar=False
    ).tolist()

def embed_query(q):
    return embed_model.encode([f"query: {q}"]).tolist()

# ── RAG core ──────────────────────────────────────────────────────────────────
def index_document(file_obj):
    if file_obj is None:
        return "⚠️ Chưa chọn file.", collection.count()
    filename = Path(file_obj.name).name
    try:
        text = load_document(file_obj.name)
    except ValueError as e:
        return f"❌ {e}", collection.count()
    chunks = split_text(text)
    if not chunks:
        return "⚠️ Tài liệu rỗng.", collection.count()
    embs  = embed_passages(chunks)
    ids   = [f"{filename}_{i}_{uuid.uuid4().hex[:6]}" for i in range(len(chunks))]
    metas = [{"source": filename, "chunk": i} for i in range(len(chunks))]
    collection.add(documents=chunks, embeddings=embs, ids=ids, metadatas=metas)
    return f"✅ {filename} — {len(chunks)} chunks đã index.", collection.count()

def retrieve(query):
    if collection.count() == 0:
        return []
    res = collection.query(
        query_embeddings=embed_query(query),
        n_results=min(TOP_K, collection.count()),
        include=["documents", "metadatas", "distances"],
    )
    return [
        {"text": d, "source": m["source"]}
        for d, m, s in zip(res["documents"][0], res["metadatas"][0], res["distances"][0])
    ]

def answer_question(question, history):
    if not question.strip():
        return history, ""
    if collection.count() == 0:
        history.append({"role": "user",      "content": question})
        history.append({"role": "assistant", "content": "⚠️ Chưa có tài liệu. Vui lòng upload trước."})
        return history, ""
    chunks = retrieve(question)
    if not chunks:
        history.append({"role": "user",      "content": question})
        history.append({"role": "assistant", "content": "Không tìm thấy thông tin liên quan."})
        return history, ""
    context = "\n\n---\n\n".join(f"[{c['source']}]\n{c['text']}" for c in chunks)
    system  = ("Bạn là trợ lý AI. Trả lời câu hỏi DỰA TRÊN ngữ cảnh. "
               "Nếu không có trong ngữ cảnh, nói rõ. Trả lời bằng tiếng Việt.")
    msgs    = [
        {"role": "system", "content": system},
        {"role": "user",   "content": f"Ngữ cảnh:\n{context}\n\nCâu hỏi: {question}"},
    ]
    prompt  = tokenizer.apply_chat_template(msgs, tokenize=False, add_generation_prompt=True)
    outputs = llm_pipeline(prompt, max_new_tokens=MAX_NEW_TOKENS,
                           do_sample=False, return_full_text=False)
    answer  = outputs[0]["generated_text"].strip()
    sources = list({c["source"] for c in chunks})
    src_text = "  \n📄 *Nguồn: " + " · ".join(sources) + "*"
    history.append({"role": "user",      "content": question})
    history.append({"role": "assistant", "content": answer + src_text})
    return history, ""

def reset_index():
    global collection
    chroma_client.delete_collection(COLLECTION_NAME)
    collection = chroma_client.get_or_create_collection(
        name=COLLECTION_NAME, metadata={"hnsw:space": "cosine"})
    return "🗑️ Đã xóa toàn bộ.", 0

def get_count():
    return collection.count()

# ── CSS ───────────────────────────────────────────────────────────────────────
CSS = """
@import url('https://fonts.googleapis.com/css2?family=DM+Sans:ital,opsz,wght@0,9..40,300;0,9..40,400;0,9..40,500;0,9..40,600;1,9..40,400&family=DM+Mono:wght@400;500&display=swap');

*, *::before, *::after { box-sizing: border-box; }

body, html {
    margin: 0; padding: 0;
    background: #0b0d12 !important;
    font-family: 'DM Sans', sans-serif !important;
}

footer, .built-with { display: none !important; }

.gradio-container {
    max-width: 100% !important;
    padding: 0 !important;
    background: #0b0d12 !important;
    min-height: 100vh;
}

/* ── Sidebar ── */
.sidebar-content {
    background: #111318 !important;
    border-right: 1px solid #1c2030 !important;
    padding: 0 !important;
}

.brand-block {
    padding: 20px 18px 16px;
    border-bottom: 1px solid #1c2030;
    margin-bottom: 8px;
}
.brand-name  { font-size: 14px; font-weight: 600; color: #f1f5f9; margin: 0; }
.brand-sub   {
    display: flex; align-items: center; gap: 6px;
    font-size: 11px; color: #4ade80; margin-top: 5px;
}
.dot {
    width: 6px; height: 6px; background: #4ade80;
    border-radius: 50%; animation: pulse 2s ease-in-out infinite; flex-shrink: 0;
}
@keyframes pulse { 0%,100%{opacity:1} 50%{opacity:.3} }

.section-label {
    font-size: 9.5px; font-weight: 600;
    letter-spacing: .12em; text-transform: uppercase;
    color: #3d4a5c; padding: 14px 18px 6px;
}

/* File upload inside sidebar */
.sidebar-content .wrap { background: transparent !important; border: none !important; padding: 4px 14px !important; }
.sidebar-content [data-testid="file-upload"],
.sidebar-content .upload-container {
    background: #0b0e14 !important;
    border: 1.5px dashed #252d3d !important;
    border-radius: 10px !important;
}

/* Buttons */
.sidebar-content button.lg {
    background: linear-gradient(135deg,#2563eb,#1e40af) !important;
    border: none !important; border-radius: 8px !important;
    color: #fff !important; font-size: 13px !important; font-weight: 500 !important;
    padding: 10px !important; width: 100% !important;
    font-family: 'DM Sans', sans-serif !important; cursor: pointer !important;
    margin-top: 6px !important;
}
.sidebar-content button.stop {
    background: transparent !important; border: 1px solid #1c2030 !important;
    border-radius: 8px !important; color: #ef4444 !important;
    font-size: 12px !important; padding: 8px !important; width: 100% !important;
    font-family: 'DM Sans', sans-serif !important; cursor: pointer !important;
    margin-top: 4px !important;
}
.sidebar-content button.stop:hover { background: #1a0d0d !important; border-color: #ef4444 !important; }

/* Number / label in sidebar */
.sidebar-content label span { color: #64748b !important; font-size: 11px !important; }
.sidebar-content input[type=number] {
    background: #0b0e14 !important; border: 1px solid #1c2030 !important;
    border-radius: 8px !important; color: #60a5fa !important;
    font-family: 'DM Mono', monospace !important; font-size: 18px !important;
    font-weight: 500 !important; padding: 8px 12px !important; width: 100% !important;
}
.sidebar-content .prose p { font-size: 11.5px !important; color: #4b5563 !important; }

/* ── Main area ── */
.main-header {
    display: flex; align-items: center; justify-content: space-between;
    padding: 14px 24px; border-bottom: 1px solid #1c2030;
    background: #0b0d12;
}
.main-header h2 {
    font-size: 15px; font-weight: 600; color: #f1f5f9;
    margin: 0; display: flex; align-items: center; gap: 8px;
}
.live-pill {
    background: #0c1f12; border: 1px solid #14532d; color: #4ade80;
    font-size: 10.5px; padding: 3px 10px; border-radius: 20px;
    display: flex; align-items: center; gap: 5px;
}

/* ── Chatbot ── */
.chatbot-wrap [data-testid="chatbot"] {
    background: transparent !important; border: none !important;
}
.chatbot-wrap .message-bubble-border { border-radius: 14px !important; }
.chatbot-wrap .message.user   > div { background: #172554 !important; border: 1px solid #1e3a8a55 !important; }
.chatbot-wrap .message.bot    > div,
.chatbot-wrap .message.assistant > div { background: #111318 !important; border: 1px solid #1c2030 !important; }
.chatbot-wrap .message p    { color: #e2e8f0 !important; font-size: 14px !important; line-height: 1.6 !important; }
.chatbot-wrap .message span { color: #94a3b8 !important; }

/* ── Suggestion chips ── */
.sug-row { gap: 8px !important; padding: 10px 24px !important; flex-wrap: wrap; }
.sug-row button {
    background: #111318 !important; border: 1px solid #1c2030 !important;
    border-radius: 10px !important; color: #64748b !important;
    font-size: 12px !important; padding: 7px 14px !important;
    font-family: 'DM Sans', sans-serif !important; transition: all .2s !important;
    white-space: nowrap !important; cursor: pointer !important;
}
.sug-row button:hover { background: #151c2e !important; border-color: #3b82f6 !important; color: #94a3b8 !important; }

/* ── Input row ── */
.input-row { padding: 12px 24px !important; border-top: 1px solid #1c2030 !important; align-items: flex-end !important; }
.input-row textarea {
    background: #111318 !important; border: 1px solid #1c2030 !important;
    border-radius: 12px !important; color: #e2e8f0 !important;
    font-size: 14px !important; padding: 12px 16px !important;
    resize: none !important; font-family: 'DM Sans', sans-serif !important;
    transition: border-color .2s !important;
}
.input-row textarea:focus { border-color: #3b82f6 !important; outline: none !important; }
.input-row textarea::placeholder { color: #2d3748 !important; }
.input-row button.primary {
    background: linear-gradient(135deg,#2563eb,#1e40af) !important;
    border: none !important; border-radius: 10px !important; color: #fff !important;
    font-size: 14px !important; font-weight: 600 !important; padding: 12px 22px !important;
    white-space: nowrap !important; height: 46px !important;
    font-family: 'DM Sans', sans-serif !important; cursor: pointer !important;
}
.input-row button.secondary {
    background: transparent !important; border: 1px solid #1c2030 !important;
    border-radius: 10px !important; color: #475569 !important;
    font-size: 13px !important; padding: 12px 14px !important;
    height: 46px !important; font-family: 'DM Sans', sans-serif !important; cursor: pointer !important;
}
"""

SUGS = [
    "💡 Tóm tắt nội dung tài liệu",
    "✨ Tìm thông tin mâu thuẫn",
    "📅 Trích xuất tất cả ngày & mốc",
]

# ── UI ────────────────────────────────────────────────────────────────────────
with gr.Blocks(title="RAG · Qwen2.5-0.5B", theme=gr.themes.Base(), css=CSS) as demo:

    # ── Sidebar ───────────────────────────────────────────────────────────────
    with gr.Sidebar(elem_classes="sidebar-content"):
        gr.HTML("""
          <div class="brand-block">
            <div class="brand-name">RAG · Qwen2.5-0.5B</div>
            <div class="brand-sub"><span class="dot"></span>Local Model Active</div>
          </div>
          <div class="section-label">Data Ingestion</div>
        """)
        file_input = gr.File(
            file_types=[".pdf", ".docx", ".txt", ".md"],
            show_label=False,
        )
        upload_btn = gr.Button("⊕  Index Document", variant="primary")
        status_vis = gr.Markdown("", elem_classes="prose")
        chunk_vis  = gr.Number(value=0, label="Chunks indexed", interactive=False)
        gr.HTML('<div class="section-label">Actions</div>')
        clear_btn  = gr.Button("🗑️  Clear All Documents", variant="stop")

    # ── Main content ──────────────────────────────────────────────────────────
    gr.HTML("""
      <div class="main-header">
        <h2>💬 Chat Interface</h2>
        <div class="live-pill"><span class="dot" style="width:5px;height:5px"></span>Running</div>
      </div>
    """)

    with gr.Column(elem_classes="chatbot-wrap"):
        chatbot = gr.Chatbot(
            height=460,
            show_label=False,
            render_markdown=True,
            type="messages",
        )

    with gr.Row(elem_classes="sug-row"):
        s1 = gr.Button(SUGS[0])
        s2 = gr.Button(SUGS[1])
        s3 = gr.Button(SUGS[2])

    with gr.Row(elem_classes="input-row"):
        question = gr.Textbox(
            placeholder="Đặt câu hỏi về tài liệu đã index...",
            show_label=False, lines=1, scale=5,
        )
        ask_btn = gr.Button("➤ Ask", variant="primary", scale=1)
        clr_btn = gr.Button("✕", variant="secondary", scale=0)

    # ── Events ────────────────────────────────────────────────────────────────
    upload_btn.click(fn=index_document, inputs=file_input,  outputs=[status_vis, chunk_vis])
    ask_btn.click(fn=answer_question,   inputs=[question, chatbot], outputs=[chatbot, question])
    question.submit(fn=answer_question, inputs=[question, chatbot], outputs=[chatbot, question])
    clear_btn.click(fn=reset_index,     outputs=[status_vis, chunk_vis])
    clr_btn.click(lambda: [], outputs=chatbot)
    s1.click(fn=lambda h: answer_question(SUGS[0][2:].strip(), h), inputs=chatbot, outputs=[chatbot, question])
    s2.click(fn=lambda h: answer_question(SUGS[1][2:].strip(), h), inputs=chatbot, outputs=[chatbot, question])
    s3.click(fn=lambda h: answer_question(SUGS[2][2:].strip(), h), inputs=chatbot, outputs=[chatbot, question])
    demo.load(fn=get_count, outputs=chunk_vis)

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