server.py

import os
import pickle
import re
import json
import asyncio
import tiktoken
from typing import List, Dict
from fastapi import FastAPI
from fastapi.responses import StreamingResponse
from pydantic import BaseModel
from dotenv import load_dotenv

from langchain_openai import OpenAIEmbeddings, ChatOpenAI
from langchain_chroma import Chroma
from sentence_transformers import CrossEncoder
from langchain_core.prompts import ChatPromptTemplate
from langchain_core.output_parsers import StrOutputParser
from langchain_core.documents import Document

load_dotenv()

app = FastAPI(title="Hanyang RAG Chatbot")

# --- 설정 및 로드 ---
DB_PATH = "./db/chroma"
BM25_PATH = "./db/bm25.pkl"
MODEL_NAME = "gpt-5-mini"

# 임베딩 & DB 로드 (large 모델도 존재함)
embeddings = OpenAIEmbeddings(model="text-embedding-3-small")
vector_store = Chroma(
    persist_directory=DB_PATH, 
    embedding_function=embeddings,
    collection_name="hanyang_rules"
)

# BM25 로드
with open(BM25_PATH, "rb") as f:
    bm25_data = pickle.load(f)
    bm25 = bm25_data["bm25"]
    doc_store = bm25_data["documents"]

# 토크나이저 설정
try:
    tokenizer = tiktoken.encoding_for_model(MODEL_NAME)
except KeyError:
    tokenizer = tiktoken.get_encoding("cl100k_base")

def tiktoken_tokenizer(text):
    tokens = tokenizer.encode(text)
    return [str(t) for t in tokens]

# Reranker 로드 (CPU 부하를 줄이기 위해 로컬/캐시 모델 사용)
reranker = CrossEncoder("BAAI/bge-reranker-v2-m3")
# LLM 설정
llm = ChatOpenAI(model=MODEL_NAME)

class ChatRequest(BaseModel):
    query: str
    history: List[Dict[str, str]] = []

async def generate_chat_stream(query: str):
    """
    단계별 진행상황(Log)과 최종 답변(Answer)을 실시간으로 Yield 합니다.
    형식: JSON String + \n
    """
    
    # 1. 검색 단계
    yield json.dumps({"type": "log", "content": "🔍 [1/4] Hybrid Search(벡터+BM25) 수행 중..."}) + "\n"
    await asyncio.sleep(0.1)

    # --- 검색 로직 ---
    # 1. Dense
    dense_results = vector_store.similarity_search_with_score(query, k=10)
    
    # 2. BM25
    tokenized_query = tiktoken_tokenizer(query)
    bm25_scores = bm25.get_scores(tokenized_query)
    top_n_bm25_indices = sorted(range(len(bm25_scores)), key=lambda i: bm25_scores[i], reverse=True)[:10]
    bm25_results = [doc_store[i] for i in top_n_bm25_indices]

    # 3. Rule-based
    rule_results = []
    match = re.search(r"제\s*(\d+)\s*조", query)
    if match:
        target_article = match.group(1)
        for doc in doc_store:
            if doc.metadata.get("article_id") == target_article:
                rule_results.append(doc)
    
    yield json.dumps({"type": "log", "content": f"📊 검색 완료: Dense({len(dense_results)}) / BM25({len(bm25_results)}) / Rule({len(rule_results)})"}) + "\n"

    # --- RRF Fusion ---
    yield json.dumps({"type": "log", "content": "🔄 [2/4] RRF 알고리즘으로 결과 통합 중..."}) + "\n"
    
    rrf_k = 60
    doc_scores = {}
    doc_obj_map = {}
    
    def update_rrf(docs, weight=1.0):
        for rank, doc in enumerate(docs):
            unique_key = doc.page_content + doc.metadata.get("source", "")
            if unique_key not in doc_scores:
                doc_scores[unique_key] = 0.0
                doc_obj_map[unique_key] = doc
            doc_scores[unique_key] += weight * (1 / (rrf_k + rank + 1))

    update_rrf([d[0] for d in dense_results])
    update_rrf(bm25_results)
    update_rrf(rule_results, weight=3.0)

    sorted_docs = sorted(doc_scores.items(), key=lambda x: x[1], reverse=True)
    # Reranking 후보군을 10개로 제한
    candidates = [doc_obj_map[key] for key, score in sorted_docs[:10]]

    # --- Reranking ---
    yield json.dumps({"type": "log", "content": f"⚖️ [3/4] Cross-Encoder 재순위화 (후보 {len(candidates)}개)..."}) + "\n"
    
    final_docs = []
    if candidates:
        pairs = []
        for doc in candidates:
            source_prefix = f"[{doc.metadata.get('source', '문서')}] "
            pairs.append([query, source_prefix + doc.page_content])
        
        # CPU 연산 병목 구간
        scores = reranker.predict(pairs)
        scored_docs = sorted(zip(candidates, scores), key=lambda x: x[1], reverse=True)
        final_docs = [doc for doc, score in scored_docs[:5]] # 최종 5개

        # 상세 로그 전송
        for i, (doc, score) in enumerate(scored_docs[:3]): # 상위 3개 점수 공개
            log_msg = f"   - Rank {i+1}: {doc.metadata.get('source')} (Score: {score:.4f})"
            yield json.dumps({"type": "log", "content": log_msg}) + "\n"
    
    # --- LLM Generation ---
    yield json.dumps({"type": "log", "content": "🤖 [4/4] GPT-5-mini 기반 답변 생성 중..."}) + "\n"

    context_text = ""
    for doc in final_docs:
        source = doc.metadata.get("source", "Unknown")
        context_text += f"📄 문서: {source}\n내용: {doc.page_content}\n\n"

    system_prompt = """당신은 한양대학교 학칙 및 규정 안내 챗봇입니다. 
    질문에 답변할 때는 반드시 근거 문서의 출처(예: 학칙 제5조, 부전공 시행세칙 등)를 언급하세요.
    Context에 없는 내용은 답변하지 마세요."""
    
    prompt = ChatPromptTemplate.from_messages([
        ("system", system_prompt),
        ("user", "Context:\n{context}\n\nQuestion: {question}")
    ])
    
    # Streaming LLM Output
    chain = prompt | llm | StrOutputParser()
    
    # LLM이 토큰을 생성할 때마다 클라이언트로 전송
    async for token in chain.astream({"context": context_text, "question": query}):
        yield json.dumps({"type": "answer", "content": token}) + "\n"

    yield json.dumps({"type": "log", "content": "✅ 답변 생성 완료."}) + "\n"
    
    # 참고 문서 정보 전송
    doc_info = [{"source": d.metadata.get("source"), "content": d.page_content[:100]+"..."} for d in final_docs]
    yield json.dumps({"type": "docs", "content": doc_info}) + "\n"


@app.post("/chat")
async def chat_endpoint(req: ChatRequest):
    return StreamingResponse(generate_chat_stream(req.query), media_type="application/x-ndjson")

if __name__ == "__main__":
    import uvicorn
    uvicorn.run(app, host="0.0.0.0", port=8000)