Cara Membuat Plagiarism Checker (Bahasa Indonesia) — step-by-step (Postgres + Qdrant + sentence-transformers + Nuxt)

Kalau kamu mau stop drama “si A nyontek si B” di kelas / tugas / kantor, ini tutorial praktis buat bikin plagiarism checker yang nyadar bahasa Indonesia, bisa cek whole text, per paragraf, per kalimat, dan menyimpan semua tugas ke dalam text bank (Postgres + Qdrant) supaya ketika kamu upload 40 tugas—mereka akan saling dicek otomatis.

---

Ringkasan arsitektur

• Frontend & API gateway: Nuxt (server script) — menerima request /check, meneruskan ke embedding/check service.
• Embedding & Search service: Python + FastAPI + sentence-transformers/paraphrase-multilingual-mpnet-base-v2 + qdrant-client — buat embedding, search, dan upsert vectors.
• Vector DB: Qdrant — menyimpan embedding (vector size = 768, distance = Cosine).
• Relational DB: PostgreSQL — menyimpan metadata dokumen (doc_id unik, title, created_at, raw_text).

---

Prasyarat (tools & versi)

• Docker + docker-compose
• Python 3.10+
• Node.js + Nuxt 3
• (Opsional) GPU jika mau cepat inference

---

1) Docker — jalankan Postgres + Qdrant cepat

Buat docker-compose.yml sederhana:

version: '3.8'
services:
  postgres:
    image: postgres:15
    environment:
      POSTGRES_USER: pguser
      POSTGRES_PASSWORD: pgpass
      POSTGRES_DB: plagiarism
    ports:
      - "5432:5432"
    volumes:
      - ./pgdata:/var/lib/postgresql/data

  qdrant:
    image: qdrant/qdrant:latest
    ports:
      - "6333:6333"
    volumes:
      - ./qdrant_storage:/qdrant/storage

Jalankan:

docker-compose up -d

---

2) Backend embedding & search (Python + FastAPI)

Buat virtualenv dan install dependency utama:

python -m venv venv
source venv/bin/activate
pip install -U pip
pip install fastapi uvicorn sentence-transformers qdrant-client sqlalchemy psycopg2-binary python-multipart numpy

Buat file service/app.py — berikut versi ringkas tapi lengkap:

# service/app.py
from fastapi import FastAPI, HTTPException
from pydantic import BaseModel
from sentence_transformers import SentenceTransformer
import numpy as np
from qdrant_client import QdrantClient
from qdrant_client.http.models import VectorParams, Distance, PointStruct
from sqlalchemy import create_engine, Column, String, Text, DateTime
from sqlalchemy.orm import declarative_base, sessionmaker
from datetime import datetime
import re
import uuid

# CONFIG
QDRANT_COLLECTION = "plagiarism_vectors"
VECTOR_SIZE = 768
THRESHOLD = 0.8
QDRANT_URL = "http://localhost:6333"
DATABASE_URL = "postgresql+psycopg2://pguser:pgpass@localhost:5432/plagiarism"

# Init
app = FastAPI(title="Plagiarism Checker Service")
model = SentenceTransformer('sentence-transformers/paraphrase-multilingual-mpnet-base-v2')  # loads on startup
qdrant = QdrantClient(url=QDRANT_URL)
engine = create_engine(DATABASE_URL)
Base = declarative_base()
SessionLocal = sessionmaker(bind=engine)

# DB model
class Document(Base):
    __tablename__ = "documents"
    doc_id = Column(String, primary_key=True, index=True)
    title = Column(String)
    text = Column(Text)
    created_at = Column(DateTime, default=datetime.utcnow)

Base.metadata.create_all(bind=engine)

# Ensure Qdrant collection exists
try:
    qdrant.recreate_collection(
        collection_name=QDRANT_COLLECTION,
        vectors_config=VectorParams(size=VECTOR_SIZE, distance=Distance.COSINE)
    )
except Exception:
    # if exists, create may fail — try get or ignore
    qdrant.create_collection(collection_name=QDRANT_COLLECTION, vectors_config=VectorParams(size=VECTOR_SIZE, distance=Distance.COSINE))

# Helpers
def split_paragraphs(text: str):
    paras = [p.strip() for p in re.split(r'\n{2,}', text) if p.strip()]
    return paras if paras else [text.strip()]

def split_sentences(paragraph: str):
    # simple rule-based sentence split (works reasonably for Indo). Improve with spaCy if needed.
    sents = re.split(r'(?<=[\.\?\!])\s+', paragraph.strip())
    return [s.strip() for s in sents if s.strip()]

def cosine_sim(a: np.ndarray, b: np.ndarray):
    den = (np.linalg.norm(a) * np.linalg.norm(b))
    if den == 0: return 0.0
    return float(np.dot(a, b) / den)

def chunk_text_all(text: str):
    """Return list of chunks with type info: full, paragraphs, sentences"""
    chunks = []
    # full
    chunks.append({"type":"full","index":0,"text":text.strip()})
    paras = split_paragraphs(text)
    for i,p in enumerate(paras):
        chunks.append({"type":"paragraph","index":i,"text":p})
        sents = split_sentences(p)
        for j,s in enumerate(sents):
            chunks.append({"type":"sentence","index":f"{i}-{j}","text":s})
    return chunks

# Request model
class CheckRequest(BaseModel):
    doc_id: str
    title: str
    text: str

@app.post("/check")
def check_and_add(req: CheckRequest):
    doc_id = req.doc_id
    title = req.title
    text = req.text

    # 1) chunk
    chunks = chunk_text_all(text)
    texts = [c["text"] for c in chunks]

    # 2) embed
    embeddings = model.encode(texts, show_progress_bar=False)
    embeddings = np.array(embeddings)  # shape (n, 768)

    # 3) search each chunk in Qdrant (exclude same doc_id results to avoid self-match)
    results = {"full": [], "paragraphs": [], "sentences": []}
    for i, c in enumerate(chunks):
        vec = embeddings[i].tolist()
        # search top 5
        hits = qdrant.search(collection_name=QDRANT_COLLECTION, query_vector=vec, limit=5, with_payload=True, with_vectors=True)
        matches = []
        for h in hits:
            payload = h.payload or {}
            source_doc = payload.get("doc_id")
            # skip matches from same doc (because this doc may already be in DB)
            if source_doc == doc_id:
                continue
            # compute exact cosine using vectors returned
            if hasattr(h, "vector") and h.vector is not None:
                sim = cosine_sim(np.array(vec), np.array(h.vector))
            else:
                # fallback to score if vector not returned
                sim = float(getattr(h, "score", 0.0))
            matches.append({
                "score": sim,
                "source_doc_id": source_doc,
                "source_text": payload.get("text"),
                "source_type": payload.get("chunk_type"),
                "source_index": payload.get("chunk_index")
            })
        # sort matches desc
        matches = sorted(matches, key=lambda x: x["score"], reverse=True)
        entry = {
            "chunk_type": c["type"],
            "chunk_index": c["index"],
            "text": c["text"],
            "top_matches": matches[:5]
        }
        if c["type"] == "full":
            results["full"].append(entry)
        elif c["type"] == "paragraph":
            results["paragraphs"].append(entry)
        else:
            results["sentences"].append(entry)

    # 4) store doc metadata in Postgres (prevent duplicate by doc_id)
    db = SessionLocal()
    existing = db.query(Document).filter(Document.doc_id==doc_id).first()
    if not existing:
        newdoc = Document(doc_id=doc_id, title=title, text=text)
        db.add(newdoc)
        db.commit()
    db.close()

    # 5) upsert all chunks to Qdrant (id uses doc_id to prevent duplicates)
    points = []
    for i, c in enumerate(chunks):
        pid = f"{doc_id}__{c['type']}__{c['index']}"
        payload = {
            "doc_id": doc_id,
            "title": title,
            "chunk_type": c["type"],
            "chunk_index": c["index"],
            "text": c["text"]
        }
        points.append(PointStruct(id=pid, vector=embeddings[i].tolist(), payload=payload))
    qdrant.upsert(collection_name=QDRANT_COLLECTION, points=points)

    # 6) build report: find any sentences/paras > THRESHOLD
    flagged = {"sentences": [], "paragraphs": []}
    for s in results["sentences"]:
        if s["top_matches"] and s["top_matches"][0]["score"] >= THRESHOLD:
            flagged["sentences"].append({
                "text": s["text"],
                "best_match": s["top_matches"][0]
            })
    for p in results["paragraphs"]:
        if p["top_matches"] and p["top_matches"][0]["score"] >= THRESHOLD:
            flagged["paragraphs"].append({
                "text": p["text"],
                "best_match": p["top_matches"][0]
            })

    return {
        "status": "ok",
        "doc_id": doc_id,
        "scores": results,
        "flagged": flagged
    }

Catatan singkat:

• Endpoint /check menerima {doc_id, title, text}.
• Melakukan check (full/paragraph/sentence), lalu insert metadata ke Postgres dan upsert semua chunk ke Qdrant.
• Self-match dikecualikan saat mencari (agar dokumen baru tidak match terhadap dirinya sendiri).
• Threshold default 0.8 (atur di THRESHOLD).

---

3) Nuxt (frontend + server route)

Kamu bisa bikin server API route di Nuxt yang meneruskan request ke FastAPI (atau langsung panggil FastAPI dari frontend, tetapi lebih aman via server route).

Contoh server/api/check.post.ts (Nuxt 3 / Nitro):

// server/api/check.post.ts
import { defineEventHandler, readBody } from 'h3'

export default defineEventHandler(async (event) => {
  const body = await readBody(event)
  // ganti URL jika service beda host/port
  const FASTAPI_URL = process.env.FASTAPI_URL || "http://localhost:8000/check"
  const res = await $fetch(FASTAPI_URL, {
    method: "POST",
    body,
    headers: { "Content-Type": "application/json" }
  })
  return res
})

Frontend form (simplified):

<template>
  <form @submit.prevent="submit">
    <input v-model="docId" placeholder="doc id (unik)"/>
    <input v-model="title" placeholder="judul"/>
    <textarea v-model="text" placeholder="tempelkan teks tugas"></textarea>
    <button>Check</button>
  </form>

  <div v-if="report">
    <h3>Flagged:</h3>
    <div v-for="s in report.flagged.sentences" :key="s.text">
      <b>Kalimat:</b> {{ s.text }} — <i>match</i>: {{ s.best_match.source_doc_id }} ({{ s.best_match.score.toFixed(3) }})
      <div>source text: {{ s.best_match.source_text }}</div>
    </div>
    <div v-for="p in report.flagged.paragraphs" :key="p.text">
      <b>Paragraf:</b> {{ p.text }} — match: {{ p.best_match.source_doc_id }} ({{ p.best_match.score.toFixed(3) }})
    </div>
  </div>
</template>

<script setup>
import { ref } from 'vue'
const docId = ref(`task-${Date.now()}`)
const title = ref('')
const text = ref('')
const report = ref(null)

async function submit(){
  report.value = null
  const res = await $fetch('/api/check', {
    method: 'POST',
    body: { doc_id: docId.value, title: title.value, text: text.value }
  })
  report.value = res
}
</script>

---

4) Testing cepat (local)

1. Jalankan FastAPI:

uvicorn service.app:app --reload --port 8000

1. Jalankan Nuxt, buka form, kirim 2 dokumen mirip → lihat flagged pada response.
2. Coba upload 40 tugas: berikan tiap tugas doc_id unik (misal: tugas-2025-09-25-001) — sistem akan menyimpan semua ke bank dan saling memeriksa.
3. Jika kamu punya textbook atau paper yang menjadi referensi anak-anak, kamu bisa memasukkannya ke dalam bank.

---

5) Tips produksi / tuning

• Threshold: 0.8 untuk indikasi kuat (sesuaikan). Paraphrase kuat mungkin 0.7–0.75.
• Top_k: 5–10 cukup untuk deteksi.
• Batching: embed batch (mis. 128 per batch) untuk performance.
• GPU: pindahkan model ke GPU (device='cuda') kalau banyak dokumen.
• Dedup: gunakan doc_id sebagai primary key; idempotensi pada /check.
• Eksklusi self-match: sudah diterapkan, pastikan saat search selalu filter payload.doc_id != doc_id.
• Tokenisasi noglitches: split sentence di atas bersifat heuristik — untuk ketepatan tinggi, gunakan tokenizer bahasa Indo (spaCy model/transformer-based segmentation).
• Privacy: simpan teks hanya jika kebijakan/privacy policy mengizinkan (kuliah / bisnis sensitif).
• Scaling: Qdrant bisa dijalankan cluster; gunakan sharding & replicas untuk ukuran besar.

---

6) Checklist deploy

• [ ] Docker compose running (Postgres + Qdrant)
• [ ] FastAPI deployed (container / VM) dan reachable dari Nuxt
• [ ] Nuxt server deployed (environment variable FASTAPI_URL)
• [ ] Backup Postgres, snapshot Qdrant
• [ ] Monitoring: latency embed time, Qdrant search time

---

Penutup (TL;DR)

• Kamu punya flow: Nuxt → FastAPI(embedding+Qdrant) → Postgres
• Checker melakukan: whole text, per paragraph, per sentence.
• Output: array skor & flagged items (kalimat/paragraf > 0.8 menampilkan teks dan referensinya).
• Saat /check juga menambahkan dokumen ke bank (unik dengan doc_id) sehingga tugas saling ter-cek.

Disclaimer

Tutorial ini dibuat oleh AI. Jika Anda menemukan error, jadikanlah sebagai progres pembelajaran. Syukur-syukur kalau bisa buat tutorial yang lebih bener.

Comments

[Budi Santoso]

ada github linknya bang?