app.py

import gradio as gr
from sentence_transformers import SentenceTransformer, CrossEncoder
from supabase import create_client
import os
from dotenv import load_dotenv
from google import genai

import pandas as pd
import time
import math

load_dotenv()

GOOGLE_API_KEY = os.getenv("GEMINI_API")
if not GOOGLE_API_KEY:
    print("⚠️ Peringatan: GOOGLE_API_KEY tidak ditemukan, Gemini akan dinonaktifkan.")
    gemini_client = None
else:
    gemini_client = genai.Client(api_key=GOOGLE_API_KEY)
    
embedder = SentenceTransformer("sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2")
reranker = CrossEncoder("cross-encoder/ms-marco-MiniLM-L-6-v2") 

print("--- Daftar Model yang Tersedia ---")

if gemini_client:
    for m in gemini_client.models.list():
        print(f"Model: {m.name} | Name: {m.display_name}")

print("----------------------------------")

# === Supabase ===
supabase_url = os.getenv("SUPABASE_URL")
supabase_key = os.getenv("SUPABASE_KEY")
supabase = create_client(supabase_url, supabase_key)

def expand_query(query: str, num_variations: int = 3) -> str:
    """
    Memperluas query menggunakan Gemini 2.5 API resmi.
    - Otomatis fallback jika API error.
    - Cache hasil untuk menghemat pemanggilan API.
    """
    if not query.strip():
        return query

    if gemini_client is None:
        return f"Kegiatan usaha yang berkaitan dengan {query}"

    prompt = f"""
    Anda adalah ahli dalam sistem pencarian KBLI (Klasifikasi Baku Lapangan Usaha Indonesia) 2020. 
    Tugas Anda adalah membuat {num_variations} variasi dari kueri berikut
    untuk meningkatkan hasil pencarian.

    Kueri pengguna: "{query}"

    Buatkan {num_variations} variasi kueri yang:
    1. Menggunakan bahasa formal atau teknis (mis. istilah industri).
    2. Menggunakan bahasa sehari-hari.
    3. Mengandung kata kunci relevan lain.

    Format keluaran HARUS seperti ini:
    Variasi 1: ...
    Variasi 2: ...
    Variasi 3: ...
    """

    try:
        # Panggil Gemini
        response = gemini_client.models.generate_content(
            # model="gemini-2.5-flash",
            # # model="gemini-robotics-er-1.5-preview",
            model="gemini-3.1-flash-lite",
            contents=prompt,
        )

        text_output = response.text.strip()
        variations = []

        for line in text_output.splitlines():
            if line.lower().startswith("variasi"):
                parts = line.split(":", 1)
                if len(parts) > 1:
                    variations.append(parts[1].strip())

        if not variations:
            print("[Gemini Warning] Tidak ada variasi ditemukan. Gunakan fallback.")
            return f"Kegiatan usaha yang berkaitan dengan {query}"

        # Gabungkan hasil
        expanded = query + ". " + " ".join(variations)
        print(f"[Gemini Expand] {query} -> {expanded}")
        return expanded

    except Exception as e:
        print(f"[Gemini Error] {e}. Menggunakan fallback lokal.")
        return f"Kegiatan usaha yang berkaitan dengan {query}"

def get_embedding(text: str):
    """
    Menghasilkan embedding vector dari teks menggunakan model SentenceTransformer.
    """
    if not text:
        return []
    
    expanded_text = expand_query(text)
    embedding = embedder.encode(expanded_text, normalize_embeddings=True).tolist()
    return embedding

# ==========================================
# ABLATION STUDY
# ==========================================

# Helper Function
def apply_sigmoid(logit):
    return 1 / (1 + math.exp(-logit))

def bm25_only(query: str, match_count: int = 50):
    """Hanya Lexical / Full-Text Search (Tanpa Vector, Tanpa Gemini, Tanpa Reranker)"""
    # Catatan: Pastikan Anda sudah membuat RPC 'search_kbli_lexical' di Supabase
    response = supabase.rpc(
        "search_kbli_lexical",
        {"query_text": query, "match_count": match_count}
    ).execute()
    
    candidates = response.data or []
    return {"results": candidates[:10]}

def dense_only(query: str, match_count: int = 50):
    """Hanya Semantic Vector (Tanpa BM25, Tanpa Gemini, Tanpa Reranker)"""
    # Gunakan query asli (tanpa expand)
    embedding_q = embedder.encode(query, normalize_embeddings=True).tolist()

    response = supabase.rpc(
        "search_kbli",
        {"query_embedding": embedding_q, "match_count": match_count}
    ).execute()
    
    candidates = response.data or []
    return {"results": candidates[:10]}

def semantic_no_gemini(query: str, match_count: int = 50):
    """Semantic Vector + Reranker (TANPA Gemini Query Expansion)"""
    expanded = query # Bypass Gemini
    embedding_q = embedder.encode(expanded, normalize_embeddings=True).tolist()

    response = supabase.rpc(
        "search_kbli",
        {"query_embedding": embedding_q, "match_count": match_count}
    ).execute()
    candidates = response.data or []

    if not candidates:
        return {"results": []}

    pairs = [(expanded, c["judul"] + " " + c["deskripsi"]) for c in candidates]

    try:
        scores = reranker.predict(pairs)
    except Exception as e:
        print("Reranker error:", e)
        return {"results": sorted(candidates, key=lambda x: x.get("similarity", 0), reverse=True)[:10]}

    rerank_vals = [float(s) for s in scores]
    rmin, rmax = min(rerank_vals), max(rerank_vals)
    
    for c, s in zip(candidates, rerank_vals):
        c["rerank_score"] = s
        if rmax - rmin > 1e-9:
            c["rerank_norm"] = (s - rmin) / (rmax - rmin)
        else:
            c["rerank_norm"] = 0.0

        sim = c.get("similarity", 0.0)
        c["hybrid_score"] = 0.6 * sim + 0.4 * c["rerank_norm"]

    candidates = sorted(candidates, key=lambda x: x["hybrid_score"], reverse=True)
    return {"results": candidates[:10]}

def hybrid_search_no_gemini(query: str, match_count: int = 50):
    """Hybrid (BM25 + Dense) + Reranker (TANPA Gemini Query Expansion)"""
    expanded = query 
    embedding_q = embedder.encode(expanded, normalize_embeddings=True).tolist()

    response = supabase.rpc(
        "search_kbli_hybrid",
        {
            "query_text": query,          
            "query_embedding": embedding_q, 
            "match_count": match_count,
            "lexical_weight": 0.4, # Diabaikan di SQL, tapi wajib dikirim
            "dense_weight": 0.6    # Diabaikan di SQL, tapi wajib dikirim
        }
    ).execute()
    
    candidates = response.data or []

    if not candidates:
        return {"results": []}

    pairs = [(query, c["judul"] + " " + c["deskripsi"]) for c in candidates]

    try:
        scores = reranker.predict(pairs)
    except Exception as e:
        print("Reranker error:", e)
        return {"results": sorted(candidates, key=lambda x: x.get("similarity", 0), reverse=True)[:10]}

    # Reranker sebagai Hakim Tunggal
    for c, s in zip(candidates, scores):
        c["rerank_score"] = float(s)
        # Gunakan Sigmoid agar nilainya menjadi probabilitas pasti
        c["final_score"] = apply_sigmoid(float(s))

    # Urutkan berdasarkan keputusan mutlak dari Reranker
    candidates = sorted(candidates, key=lambda x: x["final_score"], reverse=True)
    
    return {"results": candidates[:10]}

# ==========================================
# CORE APPS
# ==========================================

def fn_semantic(query: str, match_count: int = 50):
    expanded = expand_query(query)
    embedding_q = embedder.encode(expanded, normalize_embeddings=True).tolist()

    response = supabase.rpc(
        "search_kbli",
        {"query_embedding": embedding_q, "match_count": match_count}
    ).execute()
    candidates = response.data or []

    if not candidates:
        return {"results": []}
    
    print("=== Candidates BEFORE rerank (top 10) ===")
    for c in candidates[:10]:
        print(c.get("kode"), c.get("judul")[:80], "sim=", c.get("similarity"))

    pairs = [(expanded, c["judul"] + " " + c["deskripsi"]) for c in candidates]

    try:
        scores = reranker.predict(pairs)
    except Exception as e:
        print("Reranker error:", e)
        return {"results": sorted(candidates, key=lambda x: x.get("similarity", 0), reverse=True)[:10]}

    for c, s in zip(candidates, scores):
        c["rerank_score"] = float(s)

    print("=== Candidates AFTER rerank (top 10) ===")
    for c in candidates[:10]:
        print(c.get("kode"), c.get("judul")[:80], "sim=", c.get("similarity"), "rerank=", c.get("rerank_score"))

    rerank_vals = [c["rerank_score"] for c in candidates]
    rmin, rmax = min(rerank_vals), max(rerank_vals)
    for c in candidates:
        if rmax - rmin > 1e-9:
            c["rerank_norm"] = (c["rerank_score"] - rmin) / (rmax - rmin)
        else:
            c["rerank_norm"] = 0.0

    for c in candidates:
        sim = c.get("similarity", 0.0)
        c["hybrid_score"] = 0.6 * sim + 0.4 * c["rerank_norm"]

    candidates = sorted(candidates, key=lambda x: x["hybrid_score"], reverse=True)

    return {"results": candidates[:10]}

def hybrid_search(query: str, match_count: int = 50):
    # 1. Query Expansion
    expanded = expand_query(query)
    
    # 2. Embedding
    # Kita encode query yang sudah di-expand untuk pencarian dense
    embedding_q = embedder.encode(expanded, normalize_embeddings=True).tolist()

    # 3. Panggil Hybrid Search di Supabase
    # Kita kirimkan query ASLI untuk Lexical (agar tidak terlalu banyak noise kata), 
    # dan query EXPANDED untuk Dense (embedding).
    response = supabase.rpc(
        "search_kbli_hybrid",
        {
            "query_text": query,          # Untuk Lexical Match (tsvector)
            "query_embedding": embedding_q, # Untuk Dense Match (pgvector)
            "match_count": match_count,
            "lexical_weight": 0.2,        # Bobot Lexical (bisa disesuaikan untuk Ablation Study)
            "dense_weight": 0.8           # Bobot Dense
        }
    ).execute()
    
    candidates = response.data or []

    if not candidates:
        return {"results": []}
    
    print("=== Candidates dari Hybrid DB BEFORE rerank (top 10) ===")
    for c in candidates[:10]:
        # Tampilkan similarity yang sekarang merupakan gabungan Lexical & Dense
        print(c.get("kode"), c.get("judul")[:80], "hybrid_db_sim=", c.get("similarity"))

    # 4. Reranking dengan Cross-Encoder
    # Evaluasi kecocokan antara query asli dengan dokumen kandidat
    pairs = [(expanded, c["judul"] + " " + c["deskripsi"]) for c in candidates]

    try:
        scores = reranker.predict(pairs)
    except Exception as e:
        print("Reranker error:", e)
        return {"results": sorted(candidates, key=lambda x: x.get("similarity", 0), reverse=True)[:10]}

    # 5. Normalisasi skor Reranker & Kalkulasi Final Score
    rerank_vals = [float(s) for s in scores]
    rmin, rmax = min(rerank_vals), max(rerank_vals)
    
    for c, s in zip(candidates, rerank_vals):
        c["rerank_score"] = s
        
        # Normalisasi Min-Max
        if rmax - rmin > 1e-9:
            c["rerank_norm"] = (s - rmin) / (rmax - rmin)
        else:
            c["rerank_norm"] = 0.0

        # Skor Final Sistem Neural IR Anda (gabungan Stage 1: Hybrid Retrieval + Stage 2: Reranking)
        # Anda bisa menyesuaikan bobot ini nanti
        db_hybrid_sim = c.get("similarity", 0.0)
        c["final_score"] = (0.5 * db_hybrid_sim) + (0.5 * c["rerank_norm"])

    print("=== Candidates AFTER Cross-Encoder rerank (top 10) ===")
    # Urutkan berdasarkan final_score
    candidates = sorted(candidates, key=lambda x: x["final_score"], reverse=True)
    
    for c in candidates[:10]:
        print(c.get("kode"), c.get("judul")[:80], "final_score=", c.get("final_score"), "rerank=", c.get("rerank_score"))

    # Kembalikan 10 teratas (sesuai logika asli Anda)
    return {"results": candidates[:10]}

def search_kbli(text: str):
    if not text:
        return {"embedding": [], "results": []}

    embedding = get_embedding(text)

    response = supabase.rpc(
        "search_kbli",
        {"query_embedding": embedding, "match_count": 25}
    ).execute()

    results = response.data if response.data else []
    
    if not results:
        return "<p>Tidak ditemukan hasil.</p>"

    html = """
    <style>
        .kbli-item {
            border: 1px solid #ddd;
            border-radius: 8px;
            padding: 10px;
            margin-bottom: 8px;
            transition: background 0.2s ease;
        }
        .kbli-item:hover {
            background: #f9fafb;
        }
        .kbli-title {
            font-weight: 600;
            margin: 0;
        }
        .kbli-desc {
            font-size: 13px;
            color: #4b5563;
            margin-top: 4px;
        }
        details {
            margin-top: 16px;
            border: 1px solid #ddd;
            border-radius: 6px;
            padding: 8px;
        }
        details summary {
            cursor: pointer;
            font-weight: 600;
            color: #2563eb;
        }
        
        @media (prefers-color-scheme: dark) {
            .kbli-item { border: 1px solid #374151; }
            .kbli-item:hover { background: #1f2937; }
            .kbli-title { color: #f3f4f6; }
            .kbli-desc { color: #d1d5db; }
            .kbli-item:hover .kbli-title { color: #93c5fd; }
            .kbli-item:hover .kbli-desc { color: #e5e7eb; }
            details { border: 1px solid #374151; }
            details summary { color: #60a5fa; }
        }
    </style>
    <div>
    """

    # Top 10 == // for r in results
    top_10 = results[:10]
    for r in top_10:
        html += f"""
        <div class="kbli-item">
            <p class="kbli-title">{r['kode']} – {r['judul']}</p>
            <p class="kbli-desc">{r['deskripsi']}</p>
        </div>
        """

    # Expandable for
    others = results[10:]
    if others:
        html += "<details><summary>Lihat hasil lainnya</summary><div style='margin-top:10px;'>"
        for r in others:
            html += f"""
            <div class="kbli-item">
                <p class="kbli-title">{r['kode']} – {r['judul']}</p>
                <p class="kbli-desc">{r['deskripsi']}</p>
            </div>
            """
        html += "</div></details>"
    # End

    html += "</div>"
    return html

def calculate_mrr(retrieved_kodes, relevant_kodes_set):
    for i, kode in enumerate(retrieved_kodes):
        if kode in relevant_kodes_set:
            return 1.0 / (i + 1)
    return 0.0

def calculate_recall(retrieved_kodes, relevant_kodes_set, k=10):
    retrieved_k_set = set(retrieved_kodes[:k])

    if not relevant_kodes_set:
        return 0.0

    return len(retrieved_k_set & relevant_kodes_set) / len(relevant_kodes_set)

def calculate_ndcg(retrieved_kodes, relevance_dict, k=10):

    dcg = 0
    for i, kode in enumerate(retrieved_kodes[:k]):
        rel = relevance_dict.get(kode, 0)
        dcg += rel / math.log2(i + 2)

    ideal_rels = sorted(relevance_dict.values(), reverse=True)[:k]

    idcg = 0
    for i, rel in enumerate(ideal_rels):
        idcg += rel / math.log2(i + 2)

    return dcg / idcg if idcg > 0 else 0.0

def run_evaluation(file_obj, scenario):
    if file_obj is None:
        return "Peringatan: Silakan unggah file ground_truth.csv terlebih dahulu.", None, None
    
    df = pd.read_csv(file_obj.name)
    queries = df.groupby('query_id').first()['query'].to_dict()
    
    ground_truth = {}
    for q_id, group in df.groupby('query_id'):
        ground_truth[q_id] = dict(zip(group['kode_kbli'].astype(str), group['relevance']))

    results_list = []
    retrieval_rows = []
    
    for q_id, query_text in queries.items():
        start_time = time.perf_counter() # Mulai hitung latensi
        
        # Eksekusi fungsi berdasarkan skenario yang dipilih

        if scenario == "BM25 Only (Lexical)":
            response = bm25_only(query_text, match_count=50)
        elif scenario == "Dense Only (Semantic)":
            response = dense_only(query_text, match_count=50)
        elif scenario == "Semantic + Reranker (No Gemini)":
            response = semantic_no_gemini(query_text, match_count=50)
        elif scenario == "Semantic + Reranker (With Gemini)":
            response = fn_semantic(query_text, match_count=50)
        elif scenario == "Hybrid + Reranker (No Gemini)":
            response = hybrid_search_no_gemini(query_text, match_count=50)
        elif scenario == "Hybrid + Reranker (With Gemini)":
            response = hybrid_search(query_text, match_count=50)
        else:
            response = {"results": []}
            
        latency = time.perf_counter() - start_time # Hitung selisih waktu
        
        candidates = response.get("results", [])
        retrieved_kodes = [str(r.get('kode')) for r in candidates]

        for rank, kode in enumerate(retrieved_kodes, start=1):
            retrieval_rows.append({
                "query_id": q_id,
                "query": query_text,
                "scenario": scenario,
                "rank": rank,
                "kode_kbli": kode
            })
        
        rel_dict = ground_truth.get(q_id, {})
        relevant_kodes_set = {k for k, r in rel_dict.items() if r > 0}
        
        mrr = calculate_mrr(retrieved_kodes, relevant_kodes_set)
        recall = calculate_recall(retrieved_kodes, relevant_kodes_set, k=10)
        ndcg = calculate_ndcg(retrieved_kodes, rel_dict, k=10)
        
        results_list.append({
            "Query ID": q_id,
            "Query Text": query_text,
            "MRR@10": round(mrr, 4),
            "Recall@10": round(recall, 4),
            "nDCG@10": round(ndcg, 4),
            "Latency (sec)": round(latency, 4) # Menyimpan data latensi per kueri
        })

        if "With Gemini" in scenario:
            time.sleep(1) # Hindari rate limit Gemini API

    results_df = pd.DataFrame(results_list)
    
    # Hitung rata-rata
    summary = {
        "Skenario": scenario,
        "Total Query": len(queries),
        "Avg MRR@10": round(results_df["MRR@10"].mean(), 4),
        "Avg Recall@10": round(results_df["Recall@10"].mean(), 4),
        "Avg nDCG@10": round(results_df["nDCG@10"].mean(), 4),
        "Avg Latency (sec)": round(results_df["Latency (sec)"].mean(), 4)
    }
    
    # Export ke Excel
    safe_scenario_name = scenario.replace(" ", "_").replace("(", "").replace(")", "").replace("+", "plus")
    output_filename = f"Evaluasi_{safe_scenario_name}.xlsx"
    results_df.to_excel(output_filename, index=False)

    retrieval_df = pd.DataFrame(retrieval_rows)
    retrieval_filename = f"retrieval_results_{safe_scenario_name}.csv"
    retrieval_df.to_csv(retrieval_filename, index=False)
    
    return summary, results_df, output_filename, retrieval_filename

with gr.Blocks(css="""
    .title {font-size: 22px; font-weight: 700; color: #111827; margin-bottom: 4px;}
    .desc {font-size: 14px; color: #6b7280; margin-bottom: 16px;}
    button.gr-button {
        border-radius: 6px;
    }
    button.gr-button-primary, button.gr-button-secondary {
        border-radius: 6px;
    }
    
    .btn-row {display: flex; gap: 8px;}
    .btn-row > * {flex: 1;}
    
    .btn-row-search {display: flex; gap: 8px;}
    .btn-row-search > * {flex: none;}

    @media (max-width: 640px) {
        
        .btn-row, .btn-row-search {flex-direction: column-reverse;}
        .btn-row > button,
        .btn-row-search > button {
            width: 100% !important;
            flex: none;
        }
    }
""") as demo:
    gr.Markdown("<div class='title'>Semantic KBLI Search</div>")
    gr.Markdown("<div class='desc'>Cari kode KBLI dengan semantic search (Embedding + Matching)</div>")

    with gr.Tab("Embedding Only"):
        with gr.Row():
            with gr.Column(scale=1):
                inp1 = gr.Textbox(label="Masukkan teks")

                with gr.Row(elem_classes="btn-row"):
                    btn_clear1 = gr.Button("Clear", variant="secondary")
                    btn_submit1 = gr.Button("Submit", variant="primary")

            with gr.Column(scale=1):
                out1 = gr.JSON(label="Embedding Vector")
        
        inp1.submit(get_embedding, inp1, out1, api_name="get_embedding")
        btn_clear1.click(lambda: ("", None), None, [inp1, out1])
        btn_submit1.click(get_embedding, inp1, out1, api_name="get_embedding")
    
    with gr.Tab("Embedding Fine-tuned"):
        with gr.Row():
            with gr.Column(scale=1):
                inp2 = gr.Textbox(label="Masukkan teks")

                with gr.Row(elem_classes="btn-row"):
                    btn_clear2 = gr.Button("Clear", variant="secondary")
                    btn_submit2 = gr.Button("Submit", variant="primary")

            with gr.Column(scale=1):
                out2 = gr.JSON(label="Embedding Vector")
        
        inp2.submit(fn_semantic, inp2, out2, api_name="fn_semantic")
        btn_clear2.click(lambda: ("", None), None, [inp2, out2])
        btn_submit2.click(fn_semantic, inp2, out2, api_name="fn_semantic")

    with gr.Tab("Search KBLI"):
        inp3 = gr.Textbox(label="Masukkan teks")

        with gr.Row(elem_classes="btn-row-search"):
            btn_clear3 = gr.Button("Clear", variant="secondary")
            btn_submit3 = gr.Button("Submit", variant="primary")
        
        out3 = gr.HTML(label="Hasil Pencarian Semantic")
        
        inp3.submit(search_kbli, inp3, out3, api_name="search_kbli")
        btn_clear3.click(lambda: ("", None), None, [inp3, out3])
        btn_submit3.click(search_kbli, inp3, out3, api_name="search_kbli")

    with gr.Tab("Hybrid Search (Final)"):
        with gr.Row():
            with gr.Column(scale=1):
                inp4 = gr.Textbox(label="Masukkan teks")

                with gr.Row(elem_classes="btn-row"):
                    btn_clear4 = gr.Button("Clear", variant="secondary")
                    btn_submit4 = gr.Button("Submit", variant="primary")

            with gr.Column(scale=1):
                out4 = gr.JSON(label="Hasil Hybrid Search")
        
        inp4.submit(hybrid_search, inp4, out4, api_name="hybrid_search")
        btn_clear4.click(lambda: ("", None), None, [inp4, out4])
        btn_submit4.click(hybrid_search, inp4, out4, api_name="hybrid_search")

    with gr.Tab("Ablation Endpoints (API)"):
        gr.Markdown("### Individual Model")
        
        with gr.Row():
            with gr.Column(scale=1):
                inp5 = gr.Textbox(label="Masukkan kueri teks")

                with gr.Row(elem_classes="btn-row"):
                    btn_bm25 = gr.Button("BM25 Only", variant="primary")
                    btn_dense = gr.Button("Dense Only", variant="primary")
                
                with gr.Row(elem_classes="btn-row"):
                    btn_sem_no_gem = gr.Button("Semantic (No Gemini)", variant="primary")
                    btn_hyb_no_gem = gr.Button("Hybrid (No Gemini)", variant="primary")

                with gr.Row():
                    btn_clear5 = gr.Button("Clear", variant="secondary")

            with gr.Column(scale=1):
                out5 = gr.JSON(label="Hasil Pencarian Ablation")
        
        # Clear Button
        btn_clear5.click(lambda: ("", None), None, [inp5, out5])
        
        # Registrasi Event dan API Name
        btn_bm25.click(bm25_only, inputs=[inp5], outputs=[out5], api_name="bm25_only")
        btn_dense.click(dense_only, inputs=[inp5], outputs=[out5], api_name="dense_only")
        btn_sem_no_gem.click(semantic_no_gemini, inputs=[inp5], outputs=[out5], api_name="semantic_no_gemini")
        btn_hyb_no_gem.click(hybrid_search_no_gemini, inputs=[inp5], outputs=[out5], api_name="hybrid_search_no_gemini")
        
        inp5.submit(dense_only, inputs=[inp5], outputs=[out5])

    with gr.Tab("Ablation Study"):
        gr.Markdown("### Metrics & Latency")
        gr.Markdown("Unggah file `ground_truth.csv` Anda untuk menjalankan *batch testing* dan membandingkan skenario.")
        
        with gr.Row():
            with gr.Column(scale=1):
                eval_file = gr.File(label="Upload ground_truth.csv", file_types=[".csv"])
                eval_scenario = gr.Dropdown(
                    choices=[
                            "BM25 Only (Lexical)",
                            "Dense Only (Semantic)",
                            "Semantic + Reranker (No Gemini)",
                            "Semantic + Reranker (With Gemini)",
                            "Hybrid + Reranker (No Gemini)",
                            "Hybrid + Reranker (With Gemini)"
                        ],
                        label="Pilih Skenario Evaluasi"
                    )
                btn_run_eval = gr.Button("Jalankan Evaluasi Otomatis", variant="primary")
            
            with gr.Column(scale=1):
                eval_summary = gr.JSON(label="Ringkasan Skor Rata-rata & Latensi")
                eval_download = gr.File(label="Download Laporan (Excel)")
                eval_retrieval_download = gr.File(label="Download Retrieval Results (CSV)")
                
        eval_table = gr.Dataframe(label="Detail Per-Kueri")
        
        btn_run_eval.click(
            run_evaluation, 
            inputs=[eval_file, eval_scenario], 
            outputs=[eval_summary, eval_table, eval_download, eval_retrieval_download]
        )

if __name__ == "__main__":
    demo.queue().launch(show_error=True)