import json
import tempfile
from typing import List, Dict, Any

import gradio as gr
from transformers import pipeline

# --- Lazy init: Space başlarken modeli bir kez yükleyelim
quality_clf = pipeline("text-classification", model="snorkelai/instruction-response-quality")

def score_item(item: Dict[str, Any]) -> Dict[str, Any]:
    """Tek bir QA kaydını skorla ve quality alanını ekle."""
    q = item.get("question", "")
    a = item.get("answer", "")
    text = f"Q: {q}\nA: {a}"
    pred = quality_clf(text, truncation=True)[0]
    score = float(pred["score"])
    if score > 0.75:
        label = "high"
    elif score > 0.40:
        label = "medium"
    else:
        label = "low"
    item["quality"] = {"label": label, "score": round(score, 3)}
    return item

def improve_item(item: Dict[str, Any], target: str = "medium") -> Dict[str, Any]:
    """
    Düşük skorlara basit 'kural tabanlı' iyileştirme (LLM yok; hafif ve ücretsiz).
    İstersen buraya bir instruct LLM entegre edebiliriz.
    """
    label = item.get("quality", {}).get("label")
    if label in ("high",) or target == "none":
        return item

    q = item.get("question", "")
    a = item.get("answer", "")

    # Basit temizlikler: boşluk, büyük harf, noktalama
    q2 = q.strip()
    if q2 and q2[-1] not in "?":
        q2 += "?"
    if q2 and q2[0].islower():
        q2 = q2[0].upper() + q2[1:]

    a2 = a.strip()
    if a2 and a2[0].islower():
        a2 = a2[0].upper() + a2[1:]
    if a2 and a2[-1] not in ".!?":
        a2 += "."

    # Çok kısa cevapları minimal genişletme
    if len(a2.split()) < 5:
        a2 = a2 + " This answer has been clarified for brevity and precision."

    item["question"] = q2
    item["answer"] = a2
    return item

def process_json(
    file,
    auto_improve: bool,
    improve_threshold: str
):
    # JSON içeriğini yükle (liste veya tek obje destekler)
    data = json.load(open(file.name))
    items: List[Dict[str, Any]] = data if isinstance(data, list) else [data]

    # Skorla
    scored: List[Dict[str, Any]] = [score_item(dict(it)) for it in items]

    # İyileştirme isteğe bağlı
    if auto_improve:
        def needs_improve(lbl: str) -> bool:
            if improve_threshold == "low_only":
                return lbl == "low"
            elif improve_threshold == "low_and_medium":
                return lbl in ("low", "medium")
            return False

        improved = []
        for it in scored:
            lbl = it.get("quality", {}).get("label", "low")
            if needs_improve(lbl):
                it = improve_item(it)
                # yeniden skorlayalım ki farkı görelim
                it = score_item(it)
            improved.append(it)
        scored = improved

    # Özet tablo için küçük bir görünüm (id yoksa index)
    summary = []
    for idx, it in enumerate(scored):
        summary.append({
            "id": it.get("id", idx),
            "quality_label": it["quality"]["label"],
            "quality_score": it["quality"]["score"],
            "question_preview": (it.get("question") or "")[:120]
        })

    # İndirilebilir JSON oluştur
    tmp = tempfile.NamedTemporaryFile(delete=False, suffix=".json", mode="w")
    json.dump(scored, tmp, indent=2, ensure_ascii=False)
    tmp.flush()
    tmp.close()

    # Büyük JSON’u rahat okumak için Code kutusu
    pretty = json.dumps(scored[:50], indent=2, ensure_ascii=False)  # önizlemede ilk 50 satır
    if len(scored) > 50:
        pretty += f"\n\n// NOTE: Showing first 50 items. Download full file below."

    return summary, pretty, tmp.name

with gr.Blocks(title="Q&A Quality Evaluator", theme=gr.themes.Soft()) as demo:
    gr.Markdown("## Q&A Quality Evaluator\nUpload your JSON, score quality, and (optionally) auto-improve low items.")

    with gr.Row():
        inp_file = gr.File(file_types=[".json"], label="Upload JSON (list of objects)")

    with gr.Row():
        auto_switch = gr.Checkbox(label="Auto-improve low-quality items (light rules, no LLM)", value=False)
        improve_sel = gr.Radio(choices=["low_only", "low_and_medium", "none"], value="low_only",
                               label="Improve threshold")

    run_btn = gr.Button("Score (and Improve)")

    gr.Markdown("### Results")
    with gr.Tab("Summary Table"):
        out_table = gr.Dataframe(headers=["id", "quality_label", "quality_score", "question_preview"], wrap=True, height=400)
    with gr.Tab("Preview JSON"):
        # Büyük bir pencere: lines=30 ile rahat görüntü
        out_code = gr.Code(language="json", label="Preview (first 50 items)", interactive=False, lines=30)
    with gr.Tab("Download"):
        out_file = gr.File(label="Download full scored JSON")

    run_btn.click(
        process_json,
        inputs=[inp_file, auto_switch, improve_sel],
        outputs=[out_table, out_code, out_file]
    )

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