app.py

"""
GNN-based Medical RAG System - Main Application

Entry point cho HuggingFace Spaces deployment.

Pipeline đầy đủ:
    MedQuad Dataset
        → LlamaIndex PropertyGraphIndex
        → PyTorch Geometric graph
        → 2-layer GCN (message passing)
        → GNNHybridRetriever (α × semantic + (1-α) × structural)
        → Qwen3.5-4B GGUF via llama-cpp
        → Gradio UI
"""

# ══════════════════════════════════════════════════════════════════════════════
# BOOTSTRAP — phải chạy TRƯỚC mọi import khác
#
# Vấn đề: KHÔNG có pre-built wheel nào từ PyPI / abetlen/releases hoạt động
# trên HuggingFace Spaces vì tất cả đều dùng musl libc (Alpine build),
# trong khi HF Spaces chạy Debian (glibc).
#
# Giải pháp: Query GitHub API để tìm manylinux wheel đúng version + python tag,
# sau đó cài trực tiếp. Fallback về build từ source nếu không tìm được.
# ══════════════════════════════════════════════════════════════════════════════

import subprocess
import sys
import json
import urllib.request


def _find_manylinux_wheel(py_tag: str, version: str = "0.3.19") -> str | None:
    """
    Query GitHub API để tìm manylinux_x86_64 wheel cho đúng python tag.
    Thử repo eswarthammana/llama-cpp-wheels trước (chuyên build manylinux),
    sau đó fallback về abetlen/llama-cpp-python.
    """
    repos = [
        f"eswarthammana/llama-cpp-wheels",
        f"abetlen/llama-cpp-python",
    ]

    target_suffix = f"{py_tag}-{py_tag}-manylinux"

    for repo in repos:
        try:
            api_url = f"https://api.github.com/repos/{repo}/releases"
            req = urllib.request.Request(
                api_url,
                headers={"User-Agent": "python-bootstrap/1.0"},
            )
            with urllib.request.urlopen(req, timeout=15) as resp:
                releases = json.loads(resp.read())

            for release in releases[:10]:  # check 10 releases gần nhất
                tag = release.get("tag_name", "")
                # Tìm release có version khớp (hoặc gần nhất)
                if version not in tag and version.replace(".", "") not in tag.replace(".", ""):
                    continue
                assets = release.get("assets", [])
                for asset in assets:
                    name = asset.get("name", "")
                    if target_suffix in name and name.endswith(".whl"):
                        url = asset.get("browser_download_url", "")
                        print(f"[bootstrap] Tìm thấy: {name} ({repo})")
                        return url
        except Exception as e:
            print(f"[bootstrap] Không query được {repo}: {e}")
            continue

    return None


def _install_wheel(url: str) -> bool:
    """Cài wheel từ URL, trả về True nếu thành công."""
    result = subprocess.run(
        [
            sys.executable, "-m", "pip", "install",
            "--quiet",
            "--force-reinstall",
            "--no-deps",
            url,
        ],
        capture_output=True,
        text=True,
    )
    if result.returncode == 0:
        return True
    print(f"[bootstrap] pip lỗi: {result.stderr.strip()[:400]}")
    return False


def _clear_llama_cache():
    """Xoá module cache để import lại sạch sau khi cài."""
    for mod in list(sys.modules.keys()):
        if mod.startswith("llama_cpp"):
            del sys.modules[mod]


def _ensure_llama_cpp():
    """
    Đảm bảo llama-cpp-python hoạt động trên Debian/glibc (HF Spaces).
    Chiến lược:
      1. Nếu đã import được → done
      2. Query GitHub API tìm manylinux wheel → cài
      3. Thử fallback URLs hardcode cho các version phổ biến
      4. Build từ source (last resort)
    """
    # ── Bước 1: Thử import ────────────────────────────────────────────────────
    try:
        from llama_cpp import Llama  # noqa: F401
        print("[bootstrap] ✓ llama-cpp-python sẵn sàng (không cần cài lại).")
        return
    except (ImportError, OSError):
        pass

    v = sys.version_info
    py_tag = f"cp{v.major}{v.minor}"
    print(f"[bootstrap] llama-cpp-python chưa OK. Bắt đầu cài ({py_tag})...")

    # ── Bước 2: Query GitHub API tìm manylinux wheel ──────────────────────────
    print("[bootstrap] Đang tìm manylinux wheel qua GitHub API...")
    url = _find_manylinux_wheel(py_tag, version="0.3.19")
    if url and _install_wheel(url):
        _clear_llama_cache()
        print("[bootstrap] ✓ Cài thành công (GitHub API).")
        return

    # ── Bước 3: Thử fallback URLs hardcode ───────────────────────────────────
    # Các version đã biết có manylinux wheel từ eswarthammana
    FALLBACK_VERSIONS = ["0.3.14", "0.3.10", "0.3.4"]
    FALLBACK_REPOS = [
        "eswarthammana/llama-cpp-wheels",
        "mrzeeshanahmed/llama-cpp-python",  # repo manylinux cộng đồng
    ]

    for ver in FALLBACK_VERSIONS:
        for repo_slug in FALLBACK_REPOS:
            fallback_url = (
                f"https://github.com/{repo_slug}/releases/download/"
                f"v{ver}/"
                f"llama_cpp_python-{ver}-{py_tag}-{py_tag}-manylinux_x86_64.whl"
            )
            print(f"[bootstrap] Thử: v{ver} từ {repo_slug}...")
            if _install_wheel(fallback_url):
                _clear_llama_cache()
                print(f"[bootstrap] ✓ Cài thành công (fallback v{ver}).")
                return

    # ── Bước 4: Build từ source ───────────────────────────────────────────────
    print("[bootstrap] Tất cả pre-built wheel thất bại.")
    print("[bootstrap] Build từ source (có thể mất 5-10 phút)...")
    env = {
        **__import__("os").environ,
        "CMAKE_ARGS": "-DGGML_BLAS=OFF -DGGML_OPENMP=OFF",
        "FORCE_CMAKE": "1",
    }
    try:
        subprocess.run(
            [
                sys.executable, "-m", "pip", "install",
                "--quiet",
                "llama-cpp-python==0.3.19",
                "--no-binary", "llama-cpp-python",
            ],
            check=True,
            env=env,
        )
        _clear_llama_cache()
        print("[bootstrap] ✓ Build từ source thành công.")
    except subprocess.CalledProcessError as e:
        raise RuntimeError(
            "[bootstrap] FATAL: Không thể cài llama-cpp-python bằng bất kỳ cách nào. "
            "Kiểm tra log phía trên để biết chi tiết."
        ) from e


# ─── GỌI BOOTSTRAP ───────────────────────────────────────────────────────────
_ensure_llama_cpp()


# ══════════════════════════════════════════════════════════════════════════════
# IMPORTS CHÍNH (sau khi bootstrap đã xong)
# ══════════════════════════════════════════════════════════════════════════════

import os
import pickle
import gradio as gr
import torch
from typing import Tuple

from huggingface_hub import hf_hub_download, list_repo_files
from llama_index.embeddings.huggingface import HuggingFaceEmbedding
from llama_index.core.schema import QueryBundle

from utils.graph_builder import (
    load_healthcare_documents,
    build_llamaindex_property_graph,
    build_pyg_graph,
)
from utils.gnn_model import get_structural_embeddings
from utils.retriever import GNNHybridRetriever


# ══════════════════════════════════════════════════════════════════════════════
# CONSTANTS
# ══════════════════════════════════════════════════════════════════════════════

EMBED_MODEL_NAME = "BAAI/bge-small-en-v1.5"
NUM_DOCS         = 250
CACHE_PATH       = "rag_cache.pkl"
GNN_HIDDEN_DIM   = 256
DEFAULT_ALPHA    = 0.6
DEFAULT_TOP_K    = 5

MODEL_REPO = "Jackrong/Qwen3.5-4B-Neo-GGUF"
# MODEL_REPO = "Jackrong/Qwen3.5-4B-Claude-4.6-Opus-Reasoning-Distilledv2-GGUF"


# ══════════════════════════════════════════════════════════════════════════════
# INITIALIZATION
# ══════════════════════════════════════════════════════════════════════════════

print("\n" + "=" * 60)
print("  Initializing GNN-based Medical RAG System")
print("=" * 60)

# ── [1/5] Embedding Model ────────────────────────────────────────────────────
print("\n[1/5] Loading embedding model...")
embed_model = HuggingFaceEmbedding(
    model_name=EMBED_MODEL_NAME,
    max_length=512,
)
print(f"      ✓ {EMBED_MODEL_NAME} (384-dim)")

# ── [2/5] Build / Load Graph ─────────────────────────────────────────────────
if os.path.exists(CACHE_PATH):
    print(f"\n[2/5] Loading cached graph từ {CACHE_PATH}...")
    with open(CACHE_PATH, "rb") as f:
        cache = pickle.load(f)
    pyg_data = cache["pyg_data"]
    nodes    = cache["nodes"]
    raw_embs = cache["raw_embs"]
    print(f"      ✓ {len(nodes)} nodes, {pyg_data.edge_index.shape[1]} edges")
else:
    print(f"\n[2/5] Building knowledge graph từ MedQuad ({NUM_DOCS} docs)...")
    documents = load_healthcare_documents(num_samples=NUM_DOCS)
    index, text_nodes = build_llamaindex_property_graph(documents, embed_model)
    pyg_data, nodes, raw_embs = build_pyg_graph(
        text_nodes,
        embed_model,
        similarity_threshold=0.70,
    )
    with open(CACHE_PATH, "wb") as f:
        pickle.dump({"pyg_data": pyg_data, "nodes": nodes, "raw_embs": raw_embs}, f)
    print("      ✓ Graph cached to disk")

# ── [3/5] GNN Message Passing ────────────────────────────────────────────────
print("\n[3/5] Running GNN message passing...")
in_dim      = pyg_data.x.shape[1]
struct_embs = get_structural_embeddings(
    pyg_data,
    in_channels=in_dim,
    hidden_channels=GNN_HIDDEN_DIM,
)

# ── [4/5] Hybrid Retriever ────────────────────────────────────────────────────
print("\n[4/5] Initializing GNNHybridRetriever...")
retriever = GNNHybridRetriever(
    nodes=nodes,
    raw_embeddings=raw_embs,
    structural_embeddings=struct_embs,
    embed_model=embed_model,
    alpha=DEFAULT_ALPHA,
    top_k=DEFAULT_TOP_K,
)
print(f"      ✓ Retriever ready (α={DEFAULT_ALPHA}, top_k={DEFAULT_TOP_K})")

# ── [5/5] Load GGUF LLM ───────────────────────────────────────────────────────
print(f"\n[5/5] Loading GGUF model từ {MODEL_REPO}...")

try:
    all_files  = list(list_repo_files(MODEL_REPO))
    gguf_files = [f for f in all_files if f.endswith(".gguf")]
    preferred  = [f for f in gguf_files if "Q4_K_M" in f.upper()]
    model_filename = (preferred or gguf_files)[0]
    print(f"      Downloading: {model_filename}")
except Exception as e:
    print(f"      Không list được files ({e}), dùng tên mặc định...")
    model_filename = "qwen3.5-4b-neo-Q4_K_M.gguf"

model_path = hf_hub_download(
    repo_id=MODEL_REPO,
    filename=model_filename,
    local_dir="./models",
)

# Import lazy — sau khi bootstrap đảm bảo đúng wheel đã được cài
from llama_cpp import Llama  # noqa: E402

llm = Llama(
    model_path=model_path,
    n_ctx=2048,
    n_threads=os.cpu_count() or 4,
    n_gpu_layers=0,
    verbose=False,
)
print(f"      ✓ LLM loaded: {model_filename}")
print("\n" + "=" * 60)
print("  System ready!")
print("=" * 60 + "\n")


# ══════════════════════════════════════════════════════════════════════════════
# RAG PIPELINE
# ══════════════════════════════════════════════════════════════════════════════

def rag_pipeline(question: str, alpha: float, top_k: int) -> Tuple[str, str]:
    """Full RAG pipeline: Retrieve → Build Context → Generate."""
    if not question.strip():
        return "Vui lòng nhập câu hỏi.", ""

    retriever.alpha = float(alpha)
    retriever.top_k = int(top_k)

    query   = QueryBundle(query_str=question)
    results = retriever._retrieve(query)

    context_parts, source_lines = [], []
    for i, node_score in enumerate(results):
        context_parts.append(f"[{i+1}] {node_score.node.text}")
        q_meta = node_score.node.metadata.get("question", "N/A")
        source_lines.append(
            f"{i+1}. Hybrid Score={node_score.score:.4f} | "
            f"Related Q: {q_meta[:90]}"
        )

    context = "\n\n".join(context_parts)
    sources = "\n".join(source_lines)

    prompt = (
        "You are a knowledgeable and accurate medical assistant. "
        "Answer the question using ONLY the provided context. "
        "Be concise, clear, and factual. "
        "If the context does not contain enough information, "
        "say 'I don't have enough information to answer this accurately.'\n\n"
        f"Context:\n{context}\n\n"
        f"Question: {question}\n\n"
        "Answer:"
    )

    try:
        output = llm(
            prompt,
            max_tokens=400,
            temperature=0.1,
            top_p=0.9,
            repeat_penalty=1.1,
            stop=["Question:", "\n\nContext:", "\n\n\n"],
            echo=False,
        )
        answer = output["choices"][0]["text"].strip() or "Không thể tạo câu trả lời."
    except Exception as e:
        answer = f"Lỗi khi generate: {str(e)}"

    return answer, sources


def get_score_analysis(question: str, alpha: float) -> str:
    """Phân tích chi tiết hybrid scores."""
    if not question.strip():
        return "Nhập câu hỏi để phân tích."

    retriever.alpha = float(alpha)
    breakdown = retriever.get_score_breakdown(question, top_k=5)

    lines = ["📊 Score Breakdown (top 5 nodes):\n"]
    for item in breakdown:
        lines.append(
            f"Rank {item['rank']}: Hybrid={item['hybrid_score']:.4f} "
            f"| Semantic={item['semantic_score']:.4f} "
            f"| Structural={item['structural_score']:.4f}\n"
            f"  Preview: {item['text_preview'][:100]}\n"
        )
    return "\n".join(lines)


# ══════════════════════════════════════════════════════════════════════════════
# GRADIO UI
# ══════════════════════════════════════════════════════════════════════════════

with gr.Blocks(title="GNN-based Medical RAG", theme=gr.themes.Soft()) as demo:

    gr.Markdown("""
    # 🏥 GNN-based Medical RAG System

    Hệ thống hỏi đáp y tế sử dụng **Graph Neural Network-enhanced RAG**.

    | Component | Chi tiết |
    |-----------|----------|
    | **Dataset** | MedQuad Medical QA (HuggingFace) |
    | **Embeddings** | BAAI/bge-small-en-v1.5 (384-dim) |
    | **Graph** | Document cosine similarity graph |
    | **GNN** | 2-layer GCN – message passing enrichment |
    | **Retrieval** | α × semantic + (1-α) × structural |
    | **Generator** | Qwen3.5-4B-Neo GGUF via llama-cpp |
    """)

    with gr.Row():
        with gr.Column(scale=2):
            question_input = gr.Textbox(
                label="Medical Question",
                placeholder="e.g. What are the symptoms of Type 2 diabetes?",
                lines=3,
            )
            with gr.Accordion("⚙️ Retrieval Settings", open=False):
                alpha_slider = gr.Slider(
                    minimum=0.0, maximum=1.0, value=DEFAULT_ALPHA, step=0.05,
                    label="α — Semantic weight (1-α = Structural weight)",
                    info="Tăng α → semantic | Giảm α → GNN structural",
                )
                topk_slider = gr.Slider(
                    minimum=1, maximum=10, value=DEFAULT_TOP_K, step=1,
                    label="Top-K retrieved contexts",
                )
            with gr.Row():
                submit_btn  = gr.Button("🔍 Get Answer", variant="primary")
                analyze_btn = gr.Button("📊 Analyze Scores")

        with gr.Column(scale=3):
            answer_output   = gr.Textbox(label="Answer", lines=8)
            sources_output  = gr.Textbox(label="Retrieved Sources & Scores", lines=6)
            analysis_output = gr.Textbox(label="Score Analysis (optional)", lines=8)

    gr.Examples(
        examples=[
            ["What are the symptoms of Type 2 diabetes?"],
            ["How is hypertension treated?"],
            ["What causes Alzheimer's disease?"],
            ["What is the recommended treatment for asthma?"],
            ["How does the immune system respond to infection?"],
            ["What are the risk factors for heart disease?"],
        ],
        inputs=question_input,
        label="Example Questions",
    )

    submit_btn.click(
        fn=rag_pipeline,
        inputs=[question_input, alpha_slider, topk_slider],
        outputs=[answer_output, sources_output],
    )
    analyze_btn.click(
        fn=get_score_analysis,
        inputs=[question_input, alpha_slider],
        outputs=analysis_output,
    )


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