main.py

import joblib
import pickle
import pandas as pd
import os 
# Atasi masalah permission dengan cache lokal
import os  # ⬅️ PASTIKAN baris ini ada

# Set direktori cache ke tempat yang diizinkan Hugging Face Spaces
os.environ["HF_HOME"] = "/tmp/hf_cache"
os.environ["TRANSFORMERS_CACHE"] = "/tmp/hf_cache/transformers"
os.environ["HF_DATASETS_CACHE"] = "/tmp/hf_cache/datasets"


from sentence_transformers import SentenceTransformer
from fastapi.middleware.cors import CORSMiddleware
import re
import torch
import numpy as np
from fastapi import FastAPI, HTTPException
from fastapi.concurrency import run_in_threadpool 
from pydantic import BaseModel
from typing import List, Dict, Union, Optional
import faiss

from dotenv import load_dotenv
from supabase import create_client, Client 

load_dotenv()
SUPABASE_URL = os.environ.get("SUPABASE_URL")
SUPABASE_KEY = os.environ.get("SUPABASE_KEY")

if not SUPABASE_URL or not SUPABASE_KEY:
    raise RuntimeError("SUPABASE_URL and SUPABASE_KEY must be set in environment variables or .env file.")

try:
    supabase: Client = create_client(SUPABASE_URL, SUPABASE_KEY)
    print("Successfully connected to Supabase.")
except Exception as e:
    raise RuntimeError(f"Could not connect to Supabase: {e}")

try:
    model_sbert = SentenceTransformer('naufalihsan/indonesian-sbert-large')
    device = 'cuda' if torch.cuda.is_available() else 'cpu'
    model_sbert.to(device)

    model_lr = joblib.load('models/lr_98_24.pkl')
    loaded_index_faiss = faiss.read_index('embeddings/embeddings_model.faissindex')
    print(f"Indeks FAISS dimuat. Jumlah vektor: {loaded_index_faiss.ntotal}")
    with open('embeddings/metadata_model.pkl', "rb") as f:
        loaded_metadata_list = pickle.load(f)
    print(f"Metadata dimuat. Jumlah item metadata: {len(loaded_metadata_list)}")
    dataset = pd.read_csv('Datasets/final_dataset.csv')
    if 'hoax' in dataset.columns:
        dataset['hoax'] = dataset['hoax'].astype(int)
except FileNotFoundError as e:
    raise RuntimeError(f"Error loading model or data files: {e}")
except Exception as e:
    raise RuntimeError(f"An unexpected error occurred during model/data loading: {e}")

def check_database_schema():
    """
    Memanggil fungsi RPC di Supabase untuk mendapatkan detail tabel dan kolom.
    """
    print("Mencoba mengambil skema database dari Supabase...")
    try:
        # Panggil fungsi 'get_schema_details' yang harus sudah Anda buat di Supabase
        response = supabase.rpc('get_schema_details').execute()

        if response.data:
            print("Skema Database Berhasil Diambil:")
            current_table = None
            for row in response.data:
                table_name = row['table_name']
                column_name = row['column_name']
                data_type = row['data_type']

                # Tampilkan nama tabel sekali saja
                if table_name != current_table:
                    print(f"\n[Tabel: {table_name}]")
                    current_table = table_name

                print(f"  - Kolom: {column_name} (Tipe: {data_type})")

        else:
            print("Tidak ada data skema yang ditemukan atau terjadi error.")
            if hasattr(response, 'error') and response.error:
                print(f"Detail Error: {response.error}")

    except Exception as e:
        print(f"Terjadi kesalahan saat mengambil skema: {e}")

check_database_schema()

def normalize_text_for_counting(text: str) -> str:
    """Normalizes text for consistent counting."""
    if not isinstance(text, str):
        return ""
    text_val = text.lower()
    text_val = re.sub(r'[^a-z0-9\s]', '', text_val)
    text_val = re.sub(r'\s+', ' ', text_val).strip()
    return text_val

def embeddings(text: str) -> np.ndarray:
    text_normalized = text.lower()
    text_normalized = re.sub(r'[^A-Za-z0-9\s\[\]\(\)\{\}\<\>\.,!?;:\'\"-]', '', text_normalized)
    embedding_tensor = model_sbert.encode(text_normalized, show_progress_bar=False, device=device, convert_to_tensor=True)
    return embedding_tensor.cpu().numpy()

def pipeline_hoax(text: str) -> Dict[str, float]:
    embedding_val = embeddings(text)
    pred_proba = model_lr.predict_proba(embedding_val.reshape(1, -1))[0]
    return {
        "hoax_prob": float(pred_proba[1]),
        "valid_prob": float(pred_proba[0])
    }

def retrieval(embeddings_val: np.ndarray) -> tuple[np.ndarray, np.ndarray]: # Tipe hint diperjelas
    if embeddings_val.dtype != np.float32:
        embeddings_val = embeddings_val.astype(np.float32)
    
    embeddings_2d = embeddings_val.reshape(1, -1) if embeddings_val.ndim == 1 else embeddings_val
    
    distances_2d, indices_2d = loaded_index_faiss.search(embeddings_2d, 5)
    return indices_2d[0], distances_2d[0]


def retrieval_pipeline(text: str) -> pd.DataFrame:
    embedding_val = embeddings(text)
    idx, score = retrieval(embedding_val)
    result = dataset.iloc[idx].copy()
    result['score'] = score
    return result



app = FastAPI(
    title="TruthCheck API",
    description="API for Hoax Identification and Article Similarity Retrieval for Indonesian News."
)

app.add_middleware(
    CORSMiddleware,
    allow_origins=["*"],
    allow_credentials=True,
    allow_methods=["*"],
    allow_headers=["*"],
)


class CheckClaimRequest(BaseModel):
    claim_text: str

class SupportingArticle(BaseModel):
    title: str
    full_text_snippet: str
    url: str
    source: str
    published_year: Optional[str] = None
    hoax_status: str
    similarity_score: float

class CheckClaimResponse(BaseModel):
    claim_text: str
    hoax_probability: float
    valid_probability: float
    accuracy_percentage: float
    supporting_articles: List[SupportingArticle]

class PopularCheckItem(BaseModel):
    example_claim_title: str
    claim_text: str
    hoax_probability: float
    valid_probability: float
    accuracy_percentage: float
    supporting_articles: List[SupportingArticle]


def format_supporting_articles(retrieval_results_df: pd.DataFrame) -> List[SupportingArticle]:
    supporting_articles_list = []
    for _, row in retrieval_results_df.iterrows():
        published_year = str(row.get('Timestamp', ''))
        try:
            if pd.isna(row.get('Timestamp')):
                published_year = None
            else:
                original_ts_str = str(row['Timestamp'])
                if re.match(r'^\d{4}-\d{2}-\d{2}', original_ts_str):
                    published_year = pd.to_datetime(original_ts_str).strftime('%Y')
                elif re.match(r'^[A-Za-z]+ \d{1,2}, \d{4}', original_ts_str):
                    published_year = pd.to_datetime(original_ts_str).strftime('%Y')
                elif re.match(r'^[A-Za-z]+, \d{1,2} [A-Za-z]+ \d{4}', original_ts_str):
                    try:
                        published_year = pd.to_datetime(original_ts_str, format='%A, %d %B %Y %H:%M WIB').strftime('%Y')
                    except ValueError: 
                         published_year = pd.to_datetime(original_ts_str.split('WIB')[0].strip()).strftime('%Y') if 'WIB' in original_ts_str else original_ts_str

                else:
                     published_year = original_ts_str
        except Exception:
            published_year = str(row.get('Timestamp', '')) 

        full_text = str(row.get('FullText', '')) if not pd.isna(row.get('FullText')) else ''
        
        supporting_articles_list.append(
            SupportingArticle(
                title=str(row.get('Title', '')) if not pd.isna(row.get('Title')) else '',
                full_text_snippet=full_text[:300] + "..." if len(full_text) > 300 else full_text,
                url=str(row.get('Url', '')) if not pd.isna(row.get('Url')) else '',
                source=str(row.get('source', '')) if not pd.isna(row.get('source')) else '',
                published_year=published_year,
                hoax_status="HOAX" if row.get('hoax') == 1 else "VALID",
                similarity_score=float(row['score'])
            )
        )
    return supporting_articles_list

# --- API Endpoints ---

@app.post("/check-claim", response_model=CheckClaimResponse, summary="Periksa Klaim/Pernyataan")
async def check_claim(request: CheckClaimRequest):
    text = request.claim_text
    if not text or not text.strip():
        raise HTTPException(status_code=400, detail="Claim text cannot be empty.")

    normalized_text = normalize_text_for_counting(text)
    print(normalized_text)
    if normalized_text:
        def _sync_upsert_claim(norm_text: str, orig_text: str):
            try:
                response = supabase.rpc(
                    'upsert_popular_claim',
                    {'p_normalized_text': norm_text, 'p_original_text': orig_text}
                ).execute()

                if hasattr(response, 'error') and response.error:
                    print(f"Supabase RPC returned an error: {response.error}")
                else:
                    print(f"Supabase RPC call successful for: {norm_text[:50]}")

            except Exception as e:
                print(f"Exception during Supabase RPC call: {e}")

        try:
            await run_in_threadpool(_sync_upsert_claim, normalized_text, text)
        except Exception as e:
            print(f"Error running Supabase task in threadpool: {e}")

    # Jalankan pipeline seperti biasa
    hoax_prediction = await run_in_threadpool(pipeline_hoax, text) 
    retrieval_results_df = await run_in_threadpool(retrieval_pipeline, text)

    supporting_articles_list = format_supporting_articles(retrieval_results_df)

    return CheckClaimResponse(
        claim_text=text,
        hoax_probability=hoax_prediction["hoax_prob"],
        valid_probability=hoax_prediction["valid_prob"],
        accuracy_percentage=hoax_prediction["valid_prob"] * 100,
        supporting_articles=supporting_articles_list
    )

@app.get("/popular-checks", response_model=List[PopularCheckItem], summary="Pengecekan Terpopuler")
async def get_popular_checks(limit: int = 5): 
    popular_checks_results = []
    try:
        response = await run_in_threadpool(
            lambda: supabase.table("popular_claims")
                .select("original_claim_text, search_count") 
                .order("search_count", desc=True)
                .limit(limit)
                .execute()
        )
        
        if response.data:
            for item_data in response.data:
                original_text = item_data["original_claim_text"]
                
                example_title = original_text[:70] + "..." if len(original_text) > 70 else original_text

                hoax_prediction = await run_in_threadpool(pipeline_hoax, original_text)
                retrieval_results_df = await run_in_threadpool(retrieval_pipeline, original_text)
                
                supporting_articles_list = format_supporting_articles(retrieval_results_df)
                
                popular_checks_results.append(
                    PopularCheckItem(
                        example_claim_title=example_title,
                        claim_text=original_text,
                        hoax_probability=hoax_prediction["hoax_prob"],
                        valid_probability=hoax_prediction["valid_prob"],
                        accuracy_percentage=hoax_prediction["valid_prob"] * 100,
                        supporting_articles=supporting_articles_list
                    )
                )
        else:
            print("No popular claims found in Supabase or error in fetching.")
            if hasattr(response, 'error') and response.error:
                 print(f"Supabase fetch error: {response.error}")


    except Exception as e:
        print(f"Error fetching popular checks from Supabase: {e}")

    return popular_checks_results