惯性聚合 高效追踪和阅读你感兴趣的博客、新闻、科技资讯
阅读原文 在惯性聚合中打开

推荐订阅源

IT之家
IT之家
博客园_首页
S
SegmentFault 最新的问题
罗磊的独立博客
博客园 - 【当耐特】
OSCHINA 社区最新新闻
OSCHINA 社区最新新闻
阮一峰的网络日志
阮一峰的网络日志
D
Docker
雷峰网
雷峰网
Google DeepMind News
Google DeepMind News
博客园 - 司徒正美
V
V2EX
大猫的无限游戏
大猫的无限游戏
V
Visual Studio Blog
腾讯CDC
宝玉的分享
宝玉的分享
酷 壳 – CoolShell
酷 壳 – CoolShell
人人都是产品经理
人人都是产品经理
T
Tailwind CSS Blog
Vercel News
Vercel News
H
Help Net Security
博客园 - Franky
D
DataBreaches.Net
aimingoo的专栏
aimingoo的专栏

DEV Community

Authentication Security Deep Dive: From Brute Force to Salted Hashing (With Java Examples) Why AI Systems Don’t Fail — They Drift Spilling beans for how i learn for exam😁"Reinforcement Learning Cheat Sheet" I Replaced Chrome with Safari for AI Browser Automation. Here's What Broke (and What Finally Worked) How Python Borrows Other People's Work The $40 Architecture: Processing 1 Billion API Requests with 99.99% Uptime Vibe Coding: A Workflow Guide (From Zero to SaaS) Most webhook security guides protect the wrong side. The scary part is delivery. Headless CMS for TanStack Start: Build a Blog with Cosmic EU Age Verification App "Hacked in 2 Minutes" — What Actually Happened Comfy Cloud’s delete function does not actually remove files Running AI Models on GPU Cloud Servers: A Beginner Guide Event-driven media intelligence with AWS Step Functions and Bedrock I scored 500 AI prompts across 8 quality dimensions — here's what broke How to Call Google Gemini API from Next.js (Free Tier, No Backend Needed) The Portal Protocol: Reclaiming Human Connection in the Age of AI How to Fix Your Team's Scattered Knowledge Problem With a Self-Hosted Forum Intro to tc Cloud Functors: A Graph-First Mental Model for the Modern Cloud Designing Multi-Tenant Backends With Both Ownership and Team Access I Built a Neumorphic CSS Library with 77+ Components — Here's What I Learned PostgreSQL Performance Optimization: Why Connection Pooling Is Critical at Scale Cómo construí un SaaS multi-rubro para gestionar expensas en Argentina con FastAPI + Vue 3 🚀 I Built an Ethical Hacking Scanner Tool – Open Source Project I Replaced /usage and /context in Claude Code With a Single Statusline A Pythonic Way to Handle Emails (IMAP/SMTP) with Auto-Discovery and AI-Ready Design I Collected 8.9 Million Polymarket Price Points — Here's What I Found About How Markets Really Move EcoTrack AI — Carbon Footprint Tracker & Dashboard Everyone's Using AI. No One Agrees How. 5 self-hosted ebook managers worth trying in 2026 Building Your First AI Agent with LangChain: From Chatbot to Autonomous Assistant
RAG 시스템 실전 구축 (v26)
matias yoon · 2026-05-25 · via DEV Community

RAG 시스템 실전 구축 (v26)

개요

이 가이드는 실전에서 RAG(Retrieval-Augmented Generation) 시스템을 구축하는 데 필요한 모든 단계를 다룹니다. 개발자들은 단순한 RAG 시스템을 구현하는 것에서 벗어나 실제 운영 환경에서의 성능, 비용, 유지보수를 고려한 완전한 솔루션을 만들고자 합니다.

1. RAG 기초: 검색 → 보완 → 생성 루프

RAG 시스템은 세 가지 핵심 단계로 구성됩니다:

  1. 검색(Retrieval): 사용자의 질문과 관련된 문서 조각들을 데이터베이스에서 찾습니다.
  2. 보완(Augmentation): 검색된 문서를 프롬프트에 추가하여 생성 모델이 더 많은 컨텍스트를 갖도록 합니다.
  3. 생성(Generation): 보완된 프롬프트를 기반으로 답변을 생성합니다.
# 기본 RAG 루프 구현
class BasicRAG:
    def __init__(self, embedding_model, vector_db):
        self.embedding_model = embedding_model
        self.vector_db = vector_db

    def retrieve(self, query, top_k=5):
        query_embedding = self.embedding_model.encode([query])
        return self.vector_db.search(query_embedding, top_k)

    def generate(self, query, context):
        prompt = f"Context: {context}\n\nQuestion: {query}\n\nAnswer:"
        return self.llm.generate(prompt)

    def run(self, query):
        context = self.retrieve(query)
        return self.generate(query, context)

2. 청킹 전략: 의미적, 재귀적, 에이전트 기반

청킹 전략은 문서를 모델이 이해할 수 있는 단위로 분할하는 방법입니다.

의미적 청킹 (Semantic Chunking)

from sentence_transformers import SentenceTransformer
import numpy as np

class SemanticChunker:
    def __init__(self, model_name='all-MiniLM-L6-v2'):
        self.model = SentenceTransformer(model_name)

    def chunk(self, text, max_chunk_size=512):
        sentences = text.split('.')
        embeddings = self.model.encode(sentences)

        # 문장 간 거리 계산 후 의미 있는 단위로 분할
        chunks = []
        current_chunk = []
        current_length = 0

        for sentence, embedding in zip(sentences, embeddings):
            if current_length + len(sentence.split()) > max_chunk_size:
                chunks.append(' '.join(current_chunk))
                current_chunk = [sentence]
                current_length = len(sentence.split())
            else:
                current_chunk.append(sentence)
                current_length += len(sentence.split())

        if current_chunk:
            chunks.append(' '.join(current_chunk))
        return chunks

재귀적 청킹 (Recursive Chunking)

class RecursiveChunker:
    def __init__(self, chunk_size=512, overlap=50):
        self.chunk_size = chunk_size
        self.overlap = overlap

    def chunk(self, text):
        chunks = []
        start = 0

        while start < len(text):
            end = min(start + self.chunk_size, len(text))
            chunk = text[start:end]

            # 오버랩 추가
            if start > 0:
                overlap_start = max(0, start - self.overlap)
                chunk = text[overlap_start:end]

            chunks.append(chunk)
            start = end - self.overlap

        return chunks

3. 임베딩 모델 선택과 비교

모델 비교 테스트

import time
from sentence_transformers import SentenceTransformer
from transformers import AutoTokenizer, AutoModel
import torch

class EmbeddingBenchmark:
    def __init__(self):
        self.models = {
            'all-MiniLM-L6-v2': SentenceTransformer('all-MiniLM-L6-v2'),
            'all-mpnet-base-v2': SentenceTransformer('all-mpnet-base-v2'),
            'sentence-tiny': SentenceTransformer('sentence-tiny')
        }

    def benchmark(self, texts):
        results = {}
        for name, model in self.models.items():
            start_time = time.time()
            embeddings = model.encode(texts)
            end_time = time.time()

            results[name] = {
                'time': end_time - start_time,
                'memory': len(embeddings) * len(embeddings[0]) * 4  # float32
            }
        return results

# 사용 예시
benchmark = EmbeddingBenchmark()
test_texts = ["This is test text 1", "This is test text 2", "Another sample text"]
results = benchmark.benchmark(test_texts)

성능 기준 정리

  • all-MiniLM-L6-v2: 최적의 균형 (256차원, 50MB)
  • all-mpnet-base-v2: 높은 정확도 (768차원, 100MB)
  • sentence-tiny: 빠른 속도 (128차원, 20MB)

4. 벡터 데이터베이스 비교

Chroma vs Qdrant vs pgvector vs Milvus

# Chroma 예제
import chromadb
from chromadb.utils import embedding_functions

class ChromaRAG:
    def __init__(self):
        self.client = chromadb.Client()
        self.collection = self.client.get_or_create_collection(
            name="rag_collection",
            embedding_function=embedding_functions.DefaultEmbeddingFunction()
        )

    def add_documents(self, documents, ids):
        self.collection.add(documents=documents, ids=ids)

    def search(self, query, top_k=5):
        results = self.collection.query(
            query_texts=[query],
            n_results=top_k
        )
        return results['documents'][0]

# Qdrant 예제
from qdrant_client import QdrantClient
from qdrant_client.models import Filter, FieldCondition, MatchValue

class QdrantRAG:
    def __init__(self):
        self.client = QdrantClient(host="localhost", port=6333)
        self.collection_name = "rag_collection"

    def create_collection(self):
        self.client.recreate_collection(
            collection_name=self.collection_name,
            vectors_config={"size": 384, "distance": "Cosine"}
        )

    def search(self, query, top_k=5):
        results = self.client.search(
            collection_name=self.collection_name,
            query_vector=query,
            limit=top_k
        )
        return [hit.payload['text'] for hit in results]

# pgvector 예제
import psycopg2
from psycopg2.extras import RealDictCursor

class PGVectorRAG:
    def __init__(self, connection_string):
        self.conn = psycopg2.connect(connection_string)

    def search(self, query_embedding, top_k=5):
        with self.conn.cursor(cursor_factory=RealDictCursor) as cursor:
            cursor.execute("""
                SELECT text, 1 - (embedding <-> %s) as similarity
                FROM documents
                ORDER BY similarity DESC
                LIMIT %s
            """, (query_embedding, top_k))
            return cursor.fetchall()

성능 비교

데이터베이스 성능 메모리 사용량 설치 복잡도
Chroma 빠름 낮음 쉬움
Qdrant 빠름 중간 중간
pgvector 중간 높음 어려움
Milvus 빠름 높음 어려움

5. 전체 RAG 파이프라인 코드


python
import numpy as np
from sentence_transformers import SentenceTransformer
import chromadb
from chromadb.utils import embedding_functions
import logging

class CompleteRAGPipeline:
    def __init__(self, model_name='all-MiniLM-L6-v2'):
        self.embedding_model = SentenceTransformer(model_name)
        self.client = chromadb.Client()
        self.collection = self.client.get_or_create_collection(
            name="documents",
            embedding_function=embedding_functions.DefaultEmbeddingFunction()
        )
        self.logger = logging.getLogger(__name__)

    def setup(self, documents, ids):
        """문서 추가 및 임베딩"""
        try:
            embeddings = self.embedding_model.encode(documents)
            self.collection.add(
                documents=documents,
                embeddings=embeddings.tolist(),
                ids=ids
            )
            self.logger.info(f"Added {len(documents)} documents")
        except Exception as e:
            self.logger.error(f"Error adding documents: {e}")

    def retrieve_context(self, query, top_k=5):
        """문서 검색"""
        query_embedding = self.embedding_model.encode([query])
        results = self.collection.query(
            query_texts=[query],
            n_results=top_k,
            include=['documents', 'distances']
        )
        return results['documents'][0], results['distances'][0]

    def generate_answer(self, query, context):
        """답변 생성 (모델 사용)"""
        prompt = f"""
        사용자 질문: {query}


---

📥 **Get the full guide on Gumroad**: https://gumroad.com/l/auto ($7)