LangChain으로 만드는 실전 AI 챗봇 플랫폼 - 4개 프로젝트 통합기

from langchain.vectorstores import Chroma
from langchain.chains import RetrievalQA
 
# 문서 로드 및 분할
documents = load_documents("housing_faq.json")
chunks = text_splitter.split_documents(documents)
 
# 벡터 저장소 생성
vectorstore = Chroma.from_documents(chunks, embeddings)
 
# RAG 체인 구성
chain = RetrievalQA.from_chain_type(
    llm=llm,
    retriever=vectorstore.as_retriever(k=4),
    return_source_documents=True
)

# 가상 답변 생성
hypothetical_answer = llm.invoke(f"다음 질문에 대한 이상적인 답변을 작성하세요: {query}")
 
# 가상 답변으로 검색 (질문보다 관련 문서와 유사도가 높음)
relevant_docs = vectorstore.similarity_search(hypothetical_answer)

from sentence_transformers import CrossEncoder
 
reranker = CrossEncoder("cross-encoder/ms-marco-MiniLM-L-6-v2")
 
# 1차 검색 (빠르지만 덜 정확)
candidates = vectorstore.similarity_search(query, k=20)
 
# 2차 재정렬 (느리지만 정확)
scores = reranker.predict([(query, doc.page_content) for doc in candidates])
reranked = sorted(zip(candidates, scores), key=lambda x: x[1], reverse=True)[:4]

from langgraph.graph import StateGraph
 
def grade_documents(state):
    """검색된 문서의 관련성 평가"""
    relevant_docs = []
    for doc in state["documents"]:
        score = grader.invoke({"query": state["query"], "document": doc})
        if score == "relevant":
            relevant_docs.append(doc)
    return {"documents": relevant_docs, "needs_rewrite": len(relevant_docs) < 2}
 
def rewrite_query(state):
    """관련성 낮으면 쿼리 재작성"""
    new_query = rewriter.invoke(f"더 나은 검색을 위해 다음 질문을 재작성하세요: {state['query']}")
    return {"query": new_query}
 
# LangGraph 워크플로우
workflow = StateGraph(State)
workflow.add_node("retrieve", retrieve_documents)
workflow.add_node("grade", grade_documents)
workflow.add_node("rewrite", rewrite_query)
workflow.add_conditional_edges(
    "grade",
    lambda s: "rewrite" if s["needs_rewrite"] else "generate",
    {"rewrite": "rewrite", "generate": "generate"}
)

from neo4j import GraphDatabase
 
# PDF 사업보고서에서 엔티티 추출
entities = extract_entities(document)  # 회사, 인물, 금액 등
relations = extract_relations(document)  # 경쟁사, 자회사, 공급망 등
 
# Neo4j에 저장
with driver.session() as session:
    for entity in entities:
        session.run("MERGE (e:Entity {name: $name, type: $type})", entity)
    for rel in relations:
        session.run("""
            MATCH (a:Entity {name: $source}), (b:Entity {name: $target})
            MERGE (a)-[r:RELATION {type: $type}]->(b)
        """, rel)

class HousingService:
    _instance = None
 
    @classmethod
    def get_instance(cls):
        if cls._instance is None:
            cls._instance = cls._initialize()
        return cls._instance

llm = LLMFactory.create(
    provider="gemini",  # or "openai", "anthropic", "ollama"
    model_name="gemini-2.0-flash",
    temperature=0.7
)

# Stage 1: Frontend 빌드
FROM node:20-alpine AS frontend-builder
COPY frontend/ ./
RUN npm run build
 
# Stage 2: Python + 빌드된 Frontend
FROM python:3.11-slim
COPY --from=frontend-builder /app/dist ./unified_app/static/
CMD ["uvicorn", "unified_app.main:app", "--host", "0.0.0.0", "--port", "8080"]