RAG(Retrieval-Augmented Generation) \ucc57\ubd07\uc740 \uc678\ubd80 \uc9c0\uc2dd\ubca0\uc774\uc2a4\ub97c \ud65c\uc6a9\ud574 \ub354 \uc815\ud655\ud558\uace0 \ucd5c\uc2e0 \uc815\ubcf4\ub97c \uc81c\uacf5\ud558\ub294 AI \ucc57\ubd07\uc785\ub2c8\ub2e4. \uc774 \uac00\uc774\ub4dc\ub294 RAG \ucc57\ubd07\uc744 \uad6c\ucd95\ud558\ub294 \uacfc\uc815\uc744 \ub2e8\uacc4\ubcc4\ub85c \uc124\uba85\ud569\ub2c8\ub2e4.<\/p>\n
RAG\ub294 \uac80\uc0c9(Retrieval)<\/strong>\uacfc \uc0dd\uc131(Generation)<\/strong>\uc744 \uacb0\ud569\ud55c \uc811\uadfc \ubc29\uc2dd\uc785\ub2c8\ub2e4. \uc0ac\uc6a9\uc790 \uc9c8\ubb38\uc5d0 \ub300\ud574 \uad00\ub828 \ubb38\uc11c\ub97c \uba3c\uc800 \uac80\uc0c9\ud558\uace0, \uadf8 \uc815\ubcf4\ub97c \ubc14\ud0d5\uc73c\ub85c \ub2f5\ubcc0\uc744 \uc0dd\uc131\ud569\ub2c8\ub2e4.<\/p>\n OPENAI_API_KEY \ubc1c\uae09\ubc1b\uae30<\/a><\/p>\n \uc5ec\uae30\uc11c\ubd80\ud130 \uc544\ub798\uc758 \ubb38\uc11c\ub294 \uc778\uacf5\uc9c0\ub2a5\uc5d0 \uc758\ud574 \uc791\uc131\ub418\uc5c8\uc73c\uba70, \ud14c\uc2a4\ud2b8 \ub418\uc9c0 \uc54a\uc740 \ucf54\ub4dc\uc785\ub2c8\ub2e4.<\/b><\/font><\/p>\n RAG \ucc57\ubd07\uc740 \uae30\uc874 LLM\uc758 \ud55c\uacc4\ub97c \uadf9\ubcf5\ud558\uace0 \ub3c4\uba54\uc778 \ud2b9\ud654\ub41c \uace0\ud488\uc9c8 \ub2f5\ubcc0\uc744 \uc81c\uacf5\ud558\ub294 \uac15\ub825\ud55c \uc194\ub8e8\uc158\uc785\ub2c8\ub2e4. \uc774 \uac00\uc774\ub4dc\ub97c \ud1b5\ud574 \uae30\ubcf8\uc801\uc778 RAG \uc2dc\uc2a4\ud15c\ubd80\ud130 \uace0\uae09 \uae30\ub2a5\uae4c\uc9c0 \ub2e8\uacc4\ubcc4\ub85c \uad6c\ucd95\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n \uc131\uacf5\uc801\uc778 RAG \ucc57\ubd07 \uad6c\ucd95\uc744 \uc704\ud574\uc11c\ub294 \uc9c0\uc18d\uc801\uc778 \ubaa8\ub2c8\ud130\ub9c1\uacfc \uac1c\uc120\uc774 \ud544\uc694\ud558\uba70, \uc0ac\uc6a9\uc790 \ud53c\ub4dc\ubc31\uc744 \uc801\uadf9\uc801\uc73c\ub85c \ubc18\uc601\ud574 \uc2dc\uc2a4\ud15c\uc758 \ud488\uc9c8\uc744 \ud5a5\uc0c1\uc2dc\ucf1c\uc57c \ud569\ub2c8\ub2e4.<\/p>\n RAG \ucc57\ubd07 \uad6c\ucd95 \uac00\uc774\ub4dc \uac1c\uc694 RAG(Retrieval-Augmented Generation) \ucc57\ubd07\uc740 \uc678\ubd80 \uc9c0\uc2dd\ubca0\uc774\uc2a4\ub97c \ud65c\uc6a9\ud574 \ub354 \uc815\ud655\ud558\uace0 \ucd5c\uc2e0 \uc815\ubcf4\ub97c \uc81c\uacf5\ud558\ub294 AI \ucc57\ubd07\uc785\ub2c8\ub2e4. \uc774 \uac00\uc774\ub4dc\ub294 RAG \ucc57\ubd07\uc744 \uad6c\ucd95\ud558\ub294 \uacfc\uc815\uc744 \ub2e8\uacc4\ubcc4\ub85c \uc124\uba85\ud569\ub2c8\ub2e4. RAG\ub780 \ubb34\uc5c7\uc778\uac00? RAG\ub294 \uac80\uc0c9(Retrieval)\uacfc \uc0dd\uc131(Generation)\uc744 \uacb0\ud569\ud55c \uc811\uadfc \ubc29\uc2dd\uc785\ub2c8\ub2e4. \uc0ac\uc6a9\uc790 \uc9c8\ubb38\uc5d0 \ub300\ud574 \uad00\ub828 \ubb38\uc11c\ub97c \uba3c\uc800 \uac80\uc0c9\ud558\uace0, \uadf8 \uc815\ubcf4\ub97c \ubc14\ud0d5\uc73c\ub85c \ub2f5\ubcc0\uc744 \uc0dd\uc131\ud569\ub2c8\ub2e4. RAG\uc758 \uc7a5\uc810 \uc815\ud655\uc131: \uc2e4\uc81c \ubb38\uc11c\uc5d0 \uae30\ubc18\ud55c \ub2f5\ubcc0 \uc81c\uacf5 \ucd5c\uc2e0\uc131:\u2026 Read More »<\/a><\/span><\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[12],"tags":[],"class_list":["post-10493","post","type-post","status-publish","format-standard","hentry","category-devops"],"_links":{"self":[{"href":"https:\/\/www.skyer9.pe.kr\/wordpress\/index.php?rest_route=\/wp\/v2\/posts\/10493","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.skyer9.pe.kr\/wordpress\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.skyer9.pe.kr\/wordpress\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.skyer9.pe.kr\/wordpress\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.skyer9.pe.kr\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=10493"}],"version-history":[{"count":14,"href":"https:\/\/www.skyer9.pe.kr\/wordpress\/index.php?rest_route=\/wp\/v2\/posts\/10493\/revisions"}],"predecessor-version":[{"id":10508,"href":"https:\/\/www.skyer9.pe.kr\/wordpress\/index.php?rest_route=\/wp\/v2\/posts\/10493\/revisions\/10508"}],"wp:attachment":[{"href":"https:\/\/www.skyer9.pe.kr\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=10493"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.skyer9.pe.kr\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=10493"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.skyer9.pe.kr\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=10493"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}RAG\uc758 \uc7a5\uc810<\/h3>\n
\n
\uc2dc\uc2a4\ud15c \uc544\ud0a4\ud14d\ucc98<\/h2>\n
\uc0ac\uc6a9\uc790 \uc9c8\ubb38 \u2192 \uc784\ubca0\ub529 \u2192 \ubca1\ud130 \uac80\uc0c9 \u2192 \ubb38\uc11c \uac80\uc0c9 \u2192 \ucee8\ud14d\uc2a4\ud2b8 + \uc9c8\ubb38 \u2192 LLM \u2192 \ub2f5\ubcc0<\/code><\/pre>\n\uc8fc\uc694 \uad6c\uc131 \uc694\uc18c<\/h3>\n
\n
\ub2e8\uacc4\ubcc4 \uad6c\ucd95 \uac00\uc774\ub4dc<\/h2>\n
1\ub2e8\uacc4: \ud658\uacbd \uc124\uc815<\/h3>\n
# \ud544\uc694\ud55c \ub77c\uc774\ube0c\ub7ec\ub9ac \uc124\uce58\npip install langchain chromadb openai sentence-transformers pypdf\n\npip install -U langchain-community\npip install -U langchain-openai<\/code><\/pre>\n2\ub2e8\uacc4: \ubb38\uc11c \uc804\ucc98\ub9ac<\/h3>\n
from langchain_community.document_loaders import PyPDFLoader\nfrom langchain_community.document_loaders import TextLoader\nfrom langchain.text_splitter import RecursiveCharacterTextSplitter\n\ndef load_documents(file_path):\n """\ubb38\uc11c \ub85c\ub4dc"""\n if file_path.endswith('.pdf'):\n loader = PyPDFLoader(file_path)\n else:\n loader = TextLoader(file_path)\n return loader.load()\n\ndef split_documents(documents):\n """\ubb38\uc11c\ub97c \uccad\ud06c\ub85c \ubd84\ud560"""\n text_splitter = RecursiveCharacterTextSplitter(\n chunk_size=1000,\n chunk_overlap=200,\n length_function=len\n )\n return text_splitter.split_documents(documents)<\/code><\/pre>\n3\ub2e8\uacc4: \ubca1\ud130 \ub370\uc774\ud130\ubca0\uc774\uc2a4 \uad6c\ucd95<\/h3>\n
from langchain_community.vectorstores import Chroma\nfrom langchain_community.embeddings import OpenAIEmbeddings\n# from langchain_openai import OpenAIEmbeddings\n\ndef create_vectorstore(documents):\n """\ubca1\ud130 \uc800\uc7a5\uc18c \uc0dd\uc131"""\n embeddings = OpenAIEmbeddings()\n vectorstore = Chroma.from_documents(\n documents=documents,\n embedding=embeddings,\n persist_directory=".\/chroma_db"\n )\n return vectorstore<\/code><\/pre>\n4\ub2e8\uacc4: \uac80\uc0c9 \uc2dc\uc2a4\ud15c \uad6c\ud604<\/h3>\n
def create_retriever(vectorstore, search_type="similarity", k=5):\n """\uac80\uc0c9\uae30 \uc0dd\uc131"""\n retriever = vectorstore.as_retriever(\n search_type=search_type,\n search_kwargs={"k": k}\n )\n return retriever<\/code><\/pre>\n5\ub2e8\uacc4: RAG \uccb4\uc778 \uad6c\uc131<\/h3>\n
from langchain.chains import RetrievalQA\nfrom langchain_community.llms import OpenAI\n\ndef create_rag_chain(retriever):\n """RAG \uccb4\uc778 \uc0dd\uc131"""\n llm = OpenAI(temperature=0)\n qa_chain = RetrievalQA.from_chain_type(\n llm=llm,\n chain_type="stuff",\n retriever=retriever,\n return_source_documents=True\n )\n return qa_chain<\/code><\/pre>\n6\ub2e8\uacc4: \ucc57\ubd07 \uc778\ud130\ud398\uc774\uc2a4 \uad6c\ud604<\/h3>\n
import os\nfrom langchain_community.document_loaders import PyPDFLoader\nfrom langchain_community.document_loaders import TextLoader\nfrom langchain.text_splitter import RecursiveCharacterTextSplitter\nfrom langchain_community.vectorstores import Chroma\nfrom langchain_openai import OpenAIEmbeddings\nfrom langchain.chains import RetrievalQA\nfrom langchain_openai import OpenAI\nfrom langchain_openai import ChatOpenAI\n\nos.environ["OPENAI_API_KEY"] = "your-openai-api-key-here"\n\ndef load_documents(file_path):\n """\ubb38\uc11c \ub85c\ub4dc"""\n if file_path.endswith('.pdf'):\n loader = PyPDFLoader(file_path)\n else:\n loader = TextLoader(file_path)\n return loader.load()\n\ndef split_documents(documents):\n """\ubb38\uc11c\ub97c \uccad\ud06c\ub85c \ubd84\ud560"""\n text_splitter = RecursiveCharacterTextSplitter(\n chunk_size=1000,\n chunk_overlap=200,\n length_function=len\n )\n return text_splitter.split_documents(documents)\n\ndef create_vectorstore(documents):\n """\ubca1\ud130 \uc800\uc7a5\uc18c \uc0dd\uc131"""\n embeddings = OpenAIEmbeddings()\n vectorstore = Chroma.from_documents(\n documents=documents,\n embedding=embeddings,\n persist_directory=".\/chroma_db"\n )\n return vectorstore\n\ndef create_retriever(vectorstore, search_type="similarity", k=5):\n """\uac80\uc0c9\uae30 \uc0dd\uc131"""\n retriever = vectorstore.as_retriever(\n search_type=search_type,\n search_kwargs={"k": k}\n )\n return retriever\n\ndef create_rag_chain(retriever):\n """RAG \uccb4\uc778 \uc0dd\uc131"""\n # llm = OpenAI(temperature=0)\n llm = ChatOpenAI(\n model="gpt-4o-mini",\n temperature=0,\n max_tokens=200 # \uc751\ub2f5 \uae38\uc774 \uc81c\ud55c\n )\n qa_chain = RetrievalQA.from_chain_type(\n llm=llm,\n chain_type="stuff",\n retriever=retriever,\n return_source_documents=True\n )\n return qa_chain\n\nclass RAGChatbot:\n def __init__(self, documents_path):\n # \ubb38\uc11c \ub85c\ub4dc \ubc0f \uc804\ucc98\ub9ac\n documents = load_documents(documents_path)\n splits = split_documents(documents)\n\n # \ubca1\ud130 \uc800\uc7a5\uc18c \uc0dd\uc131\n self.vectorstore = create_vectorstore(splits)\n\n # \uac80\uc0c9\uae30 \ubc0f \uccb4\uc778 \uc0dd\uc131\n retriever = create_retriever(self.vectorstore)\n self.qa_chain = create_rag_chain(retriever)\n\n def chat(self, question):\n """\uc0ac\uc6a9\uc790 \uc9c8\ubb38\uc5d0 \ub300\ud55c \ub2f5\ubcc0 \uc0dd\uc131"""\n result = self.qa_chain.invoke({"query": question})\n return {\n "answer": result["result"],\n "sources": result["source_documents"]\n }\n\n# \uc0ac\uc6a9 \uc608\uc2dc\nchatbot = RAGChatbot(".\/law.pdf")\nresponse = chatbot.chat("\uc131\ub144\uc758 \uae30\uc900\uc774 \ub418\ub294 \ub098\uc774\uac00 \uc5b4\ub5bb\uac8c \ub418\uc9c0?")\nprint(response["answer"])<\/code><\/pre>\n\uc131\ub2a5 \ucd5c\uc801\ud654<\/h2>\n
\uccad\ud06c \ud06c\uae30 \uc870\uc815<\/h3>\n
# \ub2e4\uc591\ud55c \uccad\ud06c \ud06c\uae30 \uc2e4\ud5d8\nchunk_sizes = [500, 1000, 1500, 2000]\nfor size in chunk_sizes:\n text_splitter = RecursiveCharacterTextSplitter(\n chunk_size=size,\n chunk_overlap=size\/\/5\n )\n # \uc131\ub2a5 \ud3c9\uac00<\/code><\/pre>\n\ud558\uc774\ube0c\ub9ac\ub4dc \uac80\uc0c9<\/h3>\n
from langchain.retrievers import EnsembleRetriever\nfrom langchain.retrievers import BM25Retriever\n\ndef create_hybrid_retriever(documents, vectorstore):\n """\ud558\uc774\ube0c\ub9ac\ub4dc \uac80\uc0c9\uae30 \uc0dd\uc131"""\n # \ud0a4\uc6cc\ub4dc \uae30\ubc18 \uac80\uc0c9\n bm25_retriever = BM25Retriever.from_documents(documents)\n\n # \uc758\ubbf8 \uae30\ubc18 \uac80\uc0c9\n vector_retriever = vectorstore.as_retriever()\n\n # \uc559\uc0c1\ube14 \uac80\uc0c9\uae30\n ensemble_retriever = EnsembleRetriever(\n retrievers=[bm25_retriever, vector_retriever],\n weights=[0.5, 0.5]\n )\n return ensemble_retriever<\/code><\/pre>\n\uc7ac\ub7ad\ud0b9 \uc2dc\uc2a4\ud15c<\/h3>\n
from langchain.document_transformers import LongContextReorder\n\ndef rerank_documents(documents, query):\n """\ubb38\uc11c \uc7ac\ub7ad\ud0b9"""\n reordering = LongContextReorder()\n reordered_docs = reordering.transform_documents(documents)\n return reordered_docs<\/code><\/pre>\n\uace0\uae09 \uae30\ub2a5<\/h2>\n
\ub300\ud654 \uba54\ubaa8\ub9ac \ucd94\uac00<\/h3>\n
from langchain.memory import ConversationBufferMemory\nfrom langchain.chains import ConversationalRetrievalChain\n\ndef create_conversational_chain(retriever):\n """\ub300\ud654\ud615 RAG \uccb4\uc778 \uc0dd\uc131"""\n memory = ConversationBufferMemory(\n memory_key="chat_history",\n return_messages=True\n )\n\n qa_chain = ConversationalRetrievalChain.from_llm(\n llm=OpenAI(temperature=0),\n retriever=retriever,\n memory=memory\n )\n return qa_chain<\/code><\/pre>\n\uc2e4\uc2dc\uac04 \ubb38\uc11c \uc5c5\ub370\uc774\ud2b8<\/h3>\n
def update_vectorstore(vectorstore, new_documents):\n """\ubca1\ud130 \uc800\uc7a5\uc18c \uc5c5\ub370\uc774\ud2b8"""\n splits = split_documents(new_documents)\n vectorstore.add_documents(splits)\n vectorstore.persist()<\/code><\/pre>\n\ub2f5\ubcc0 \ud488\uc9c8 \ud3c9\uac00<\/h3>\n
from langchain.evaluation.qa import QAEvalChain\n\ndef evaluate_rag_system(qa_chain, test_questions):\n """RAG \uc2dc\uc2a4\ud15c \ud3c9\uac00"""\n eval_chain = QAEvalChain.from_llm(OpenAI(temperature=0))\n\n results = []\n for question in test_questions:\n result = qa_chain({"query": question})\n evaluation = eval_chain.evaluate(\n question,\n result["result"],\n result["source_documents"]\n )\n results.append(evaluation)\n\n return results<\/code><\/pre>\n\ubc30\ud3ec \ubc0f \ubaa8\ub2c8\ud130\ub9c1<\/h2>\n
FastAPI\ub97c \uc774\uc6a9\ud55c API \uc11c\ubc84<\/h3>\n
from fastapi import FastAPI\nfrom pydantic import BaseModel\n\napp = FastAPI()\nchatbot = RAGChatbot(".\/documents\/")\n\nclass ChatRequest(BaseModel):\n question: str\n\n@app.post("\/chat")\nasync def chat_endpoint(request: ChatRequest):\n response = chatbot.chat(request.question)\n return response<\/code><\/pre>\n\ub85c\uae45 \ubc0f \ubaa8\ub2c8\ud130\ub9c1<\/h3>\n
import logging\nfrom datetime import datetime\n\ndef setup_logging():\n logging.basicConfig(\n level=logging.INFO,\n format='%(asctime)s - %(levelname)s - %(message)s',\n handlers=[\n logging.FileHandler('rag_chatbot.log'),\n logging.StreamHandler()\n ]\n )\n\ndef log_interaction(question, answer, sources):\n """\uc0ac\uc6a9\uc790 \uc0c1\ud638\uc791\uc6a9 \ub85c\uadf8"""\n logging.info(f"Question: {question}")\n logging.info(f"Answer: {answer}")\n logging.info(f"Sources: {len(sources)} documents")<\/code><\/pre>\n\uc8fc\uc758\uc0ac\ud56d \ubc0f \ubaa8\ubc94 \uc0ac\ub840<\/h2>\n
\ubcf4\uc548 \uace0\ub824\uc0ac\ud56d<\/h3>\n
\n
\uc131\ub2a5 \ucd5c\uc801\ud654<\/h3>\n
\n
\ub370\uc774\ud130 \ud488\uc9c8 \uad00\ub9ac<\/h3>\n
\n
\uacb0\ub860<\/h2>\n
\ucc38\uace0 \uc790\ub8cc<\/h2>\n
\n