Initial commit: import from sinmb79/Gov-chat-bot

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

backend/app/services/audit.py

@@ -0,0 +1,39 @@
+"""
+감사 로그 기록 헬퍼.
+표준 action: faq.create|faq.update|faq.delete
+             doc.upload|doc.approve|doc.delete
+             user.restrict|user.unblock
+             crawler.approve|crawler.reject
+             config.update
+"""
+from typing import Optional
+
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from app.models.audit import AuditLog
+
+
+async def log_action(
+    db: AsyncSession,
+    tenant_id: str,
+    actor_id: str,
+    actor_type: str,  # 'admin_user' | 'system_admin'
+    action: str,
+    target_type: Optional[str] = None,
+    target_id: Optional[str] = None,
+    diff: Optional[dict] = None,
+    ip_address: Optional[str] = None,
+) -> AuditLog:
+    entry = AuditLog(
+        tenant_id=tenant_id,
+        actor_id=actor_id,
+        actor_type=actor_type,
+        action=action,
+        target_type=target_type,
+        target_id=target_id,
+        diff=diff,
+        ip_address=ip_address,
+    )
+    db.add(entry)
+    await db.commit()
+    return entry

backend/app/services/complaint_logger.py

@@ -0,0 +1,44 @@
+"""
+민원 이력 DB 저장.
+개인정보 원칙: utterance_masked(마스킹), user_key(SHA-256 해시 16자리).
+원문 미저장.
+"""
+from typing import Optional
+
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from app.models.complaint import ComplaintLog
+from app.services.masking import mask_text, hash_user_key
+from app.services.routing import RoutingResult
+
+
+async def log_complaint(
+    db: AsyncSession,
+    tenant_id: str,
+    raw_utterance: str,
+    raw_user_id: str,
+    result: RoutingResult,
+    channel: str = "kakao",
+) -> ComplaintLog:
+    """
+    민원 이력을 ComplaintLog에 기록.
+    - utterance: 마스킹 후 저장
+    - user_key: SHA-256 해시 16자리
+    - 원문 미저장
+    """
+    entry = ComplaintLog(
+        tenant_id=tenant_id,
+        user_key=hash_user_key(raw_user_id),
+        utterance_masked=mask_text(raw_utterance)[:1000],
+        channel=channel,
+        request_id=result.request_id,
+        response_tier=result.tier,
+        response_source=result.source,
+        faq_id=result.faq_id,
+        doc_id=result.doc_id,
+        response_ms=result.elapsed_ms,
+        is_timeout=result.is_timeout,
+    )
+    db.add(entry)
+    await db.commit()
+    return entry

backend/app/services/crawler.py

@@ -0,0 +1,92 @@
+"""
+웹 크롤러 — httpx + BeautifulSoup4.
+robots.txt 준수. CrawlerURL 기반.
+"""
+from typing import Optional
+from datetime import datetime, timezone
+
+import httpx
+from bs4 import BeautifulSoup
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from app.models.knowledge import CrawlerURL, Document
+
+
+CRAWLER_HEADERS = {
+    "User-Agent": "SmartBot-KR/1.0 (+https://github.com/sinmb79/Gov-chat-bot)",
+}
+CRAWL_TIMEOUT = 15  # 초
+
+
+async def check_robots_txt(base_url: str, target_path: str) -> bool:
+    """robots.txt 확인. 크롤링 허용 여부 반환."""
+    try:
+        from urllib.parse import urlparse
+        parsed = urlparse(base_url)
+        robots_url = f"{parsed.scheme}://{parsed.netloc}/robots.txt"
+        async with httpx.AsyncClient(timeout=5) as client:
+            resp = await client.get(robots_url, headers=CRAWLER_HEADERS)
+            if resp.status_code != 200:
+                return True  # robots.txt 없으면 허용으로 간주
+            content = resp.text.lower()
+            # 간단한 User-agent: * Disallow 체크
+            lines = content.splitlines()
+            in_block = False
+            for line in lines:
+                line = line.strip()
+                if line.startswith("user-agent:"):
+                    agent = line.split(":", 1)[1].strip()
+                    in_block = agent in ("*", "smartbot-kr")
+                elif in_block and line.startswith("disallow:"):
+                    disallowed = line.split(":", 1)[1].strip()
+                    if disallowed and target_path.startswith(disallowed):
+                        return False
+            return True
+    except Exception:
+        return True  # 확인 불가 시 허용
+
+
+async def crawl_url(url: str) -> Optional[str]:
+    """URL 크롤링 → 텍스트 추출. 실패 시 None."""
+    try:
+        from urllib.parse import urlparse
+        parsed = urlparse(url)
+        target_path = parsed.path or "/"
+
+        if not await check_robots_txt(url, target_path):
+            return None  # robots.txt 불허
+
+        async with httpx.AsyncClient(
+            timeout=CRAWL_TIMEOUT,
+            follow_redirects=True,
+            headers=CRAWLER_HEADERS,
+        ) as client:
+            resp = await client.get(url)
+            resp.raise_for_status()
+
+        content_type = resp.headers.get("content-type", "")
+        if "html" in content_type:
+            soup = BeautifulSoup(resp.content, "html.parser")
+            for tag in soup(["script", "style", "nav", "footer", "header"]):
+                tag.decompose()
+            return soup.get_text(separator="\n", strip=True)
+        else:
+            return resp.text
+
+    except Exception:
+        return None
+
+
+class CrawlerService:
+    def __init__(self, db: AsyncSession):
+        self.db = db
+
+    async def run(self, crawler_url: CrawlerURL, tenant_id: str) -> Optional[str]:
+        """크롤러 URL 실행 → 텍스트 반환."""
+        text = await crawl_url(crawler_url.url)
+
+        # last_crawled 업데이트
+        crawler_url.last_crawled = datetime.now(timezone.utc)
+        await self.db.commit()
+
+        return text

backend/app/services/document_processor.py

@@ -0,0 +1,88 @@
+"""
+문서 처리 파이프라인:
+파싱 → 청킹 → 임베딩 → VectorDB 저장 → Document 레코드 업데이트
+"""
+from typing import Optional
+
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from app.models.knowledge import Document
+from app.providers.embedding import EmbeddingProvider
+from app.providers.vectordb import VectorDBProvider
+from app.services.parsers.text_parser import extract_text, chunk_text
+
+
+class DocumentProcessor:
+    def __init__(
+        self,
+        embedding_provider: EmbeddingProvider,
+        vectordb_provider: VectorDBProvider,
+        db: AsyncSession,
+    ):
+        self.embedding = embedding_provider
+        self.vectordb = vectordb_provider
+        self.db = db
+
+    async def process(self, tenant_id: str, doc: Document, content: bytes) -> int:
+        """
+        문서를 파싱·청킹·임베딩하여 VectorDB에 저장.
+        chunk_count 반환. 실패 시 0.
+        """
+        # 1. 텍스트 추출
+        text = extract_text(content, doc.filename)
+        if not text or not text.strip():
+            doc.status = "parse_failed"
+            await self.db.commit()
+            return 0
+
+        # 2. 청킹
+        chunks = chunk_text(text)
+        if not chunks:
+            doc.status = "parse_failed"
+            await self.db.commit()
+            return 0
+
+        # 3. 임베딩
+        try:
+            embeddings = await self.embedding.embed(chunks)
+        except NotImplementedError:
+            doc.status = "embedding_unavailable"
+            await self.db.commit()
+            return 0
+        except Exception:
+            doc.status = "embedding_failed"
+            await self.db.commit()
+            return 0
+
+        # 4. 메타데이터 구성
+        published = doc.published_at.strftime("%Y.%m") if doc.published_at else ""
+        metadatas = [
+            {
+                "doc_id": doc.id,
+                "filename": doc.filename,
+                "chunk_idx": i,
+                "published_at": published,
+                "tenant_id": tenant_id,
+            }
+            for i in range(len(chunks))
+        ]
+
+        # 5. VectorDB 저장
+        await self.vectordb.upsert(
+            tenant_id=tenant_id,
+            doc_id=doc.id,
+            chunks=chunks,
+            embeddings=embeddings,
+            metadatas=metadatas,
+        )
+
+        # 6. Document 레코드 업데이트
+        doc.chunk_count = len(chunks)
+        doc.status = "processed"
+        await self.db.commit()
+
+        return len(chunks)
+
+    async def delete(self, tenant_id: str, doc_id: str) -> None:
+        """VectorDB에서 문서 청크 삭제."""
+        await self.vectordb.delete(tenant_id=tenant_id, doc_id=doc_id)

backend/app/services/faq_search.py

@@ -0,0 +1,94 @@
+from typing import Optional
+
+from sqlalchemy import select
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from app.models.knowledge import FAQ
+from app.providers.embedding import EmbeddingProvider
+from app.providers.vectordb import VectorDBProvider
+from app.providers.base import SearchResult
+
+FAQ_SIMILARITY_THRESHOLD = 0.85  # Tier A 기준
+
+
+class FAQSearchService:
+    """
+    Tier A — FAQ 임베딩 유사도 검색.
+    임베딩 유사도 ≥ 0.85 시 등록 FAQ 반환.
+    """
+
+    def __init__(
+        self,
+        embedding_provider: EmbeddingProvider,
+        vectordb_provider: VectorDBProvider,
+        db: AsyncSession,
+    ):
+        self.embedding = embedding_provider
+        self.vectordb = vectordb_provider
+        self.db = db
+
+    async def search(
+        self, tenant_id: str, utterance: str
+    ) -> Optional[tuple[FAQ, float]]:
+        """
+        발화를 임베딩 → 벡터DB 검색 → 0.85 이상이면 FAQ 반환.
+        없으면 None.
+        """
+        try:
+            vecs = await self.embedding.embed([utterance])
+        except NotImplementedError:
+            return None
+
+        query_vec = vecs[0]
+        results = await self.vectordb.search(
+            tenant_id=tenant_id,
+            query_vec=query_vec,
+            top_k=1,
+            threshold=FAQ_SIMILARITY_THRESHOLD,
+        )
+
+        if not results:
+            return None
+
+        top: SearchResult = results[0]
+        faq_id = top.metadata.get("faq_id")
+        if not faq_id:
+            return None
+
+        faq = await self._load_faq(tenant_id, faq_id)
+        if faq is None:
+            return None
+
+        return faq, top.score
+
+    async def _load_faq(self, tenant_id: str, faq_id: str) -> Optional[FAQ]:
+        result = await self.db.execute(
+            select(FAQ).where(
+                FAQ.tenant_id == tenant_id,
+                FAQ.id == faq_id,
+                FAQ.is_active.is_(True),
+            )
+        )
+        return result.scalar_one_or_none()
+
+    async def increment_hit(self, faq_id: str) -> None:
+        """FAQ hit_count 증가."""
+        faq = await self.db.get(FAQ, faq_id)
+        if faq:
+            faq.hit_count = (faq.hit_count or 0) + 1
+            await self.db.commit()
+
+    async def index_faq(self, tenant_id: str, faq: FAQ) -> None:
+        """FAQ를 벡터DB에 색인."""
+        text = f"{faq.question}\n{faq.answer}"
+        try:
+            vecs = await self.embedding.embed([text])
+        except NotImplementedError:
+            return
+        await self.vectordb.upsert(
+            tenant_id=tenant_id,
+            doc_id=faq.id,
+            chunks=[text],
+            embeddings=vecs,
+            metadatas=[{"faq_id": faq.id, "question": faq.question}],
+        )

backend/app/services/idempotency.py

@@ -0,0 +1,28 @@
+import json
+from typing import Optional
+
+
+class IdempotencyCache:
+    def __init__(self, redis_client):
+        self.redis = redis_client
+        self.ttl = 60  # 기본 TTL (Phase 0B에서 settings로 교체 예정)
+
+    def _key(self, tenant_id: str, request_id: str) -> str:
+        return f"idempotency:{tenant_id}:{request_id}"
+
+    async def get(self, tenant_id: str, request_id: Optional[str]) -> Optional[dict]:
+        if request_id is None:
+            return None
+        raw = await self.redis.get(self._key(tenant_id, request_id))
+        if raw is None:
+            return None
+        return json.loads(raw)
+
+    async def set(self, tenant_id: str, request_id: Optional[str], result_dict: dict) -> None:
+        if request_id is None:
+            return
+        await self.redis.setex(
+            self._key(tenant_id, request_id),
+            self.ttl,
+            json.dumps(result_dict),
+        )

backend/app/services/masking.py

@@ -0,0 +1,34 @@
+import hashlib
+import re
+
+# 마스킹 패턴 정의
+_PATTERNS = [
+    # 주민등록번호: 6자리-1~4자리+6자리 (숫자 금액과 구분: 앞에 비숫자 또는 시작, 뒤에 비숫자 또는 끝)
+    (re.compile(r"(?<!\d)\d{6}-[1-4]\d{6}(?!\d)"), "######-*######"),
+    # 전화번호: 010-1234-5678 형식
+    (re.compile(r"0\d{1,2}-\d{3,4}-\d{4}"), "***-****-****"),
+    # 이메일
+    (re.compile(r"[\w._%+\-]+@[\w.\-]+\.[a-zA-Z]{2,}"), "***@***.***"),
+    # 카드번호: 4자리씩 4그룹
+    (re.compile(r"\d{4}[- ]?\d{4}[- ]?\d{4}[- ]?\d{4}"), "****-****-****-****"),
+]
+
+
+def mask_text(text: str) -> str:
+    """텍스트에서 개인정보 패턴을 마스킹하여 반환."""
+    for pattern, replacement in _PATTERNS:
+        text = pattern.sub(replacement, text)
+    return text
+
+
+def hash_user_key(kakao_id: str) -> str:
+    """SHA-256 해시 앞 16자리 반환."""
+    return hashlib.sha256(kakao_id.encode()).hexdigest()[:16]
+
+
+def has_sensitive_data(text: str) -> bool:
+    """텍스트에 개인정보 패턴이 포함되어 있으면 True."""
+    for pattern, _ in _PATTERNS:
+        if pattern.search(text):
+            return True
+    return False

backend/app/services/metrics.py

@@ -0,0 +1,95 @@
+import json
+from typing import Optional
+
+from app.services.routing import RoutingResult
+
+METRIC_KEYS = [
+    "total_count",
+    "faq_hit_count",
+    "rag_hit_count",
+    "llm_hit_count",
+    "fallback_count",
+    "timeout_count",
+    "response_ms_sum",
+    "blocked_attempts",
+]
+
+P95_SORTED_SET = "response_ms_p95_buf"
+P95_MAX_SIZE = 10000
+
+_SOURCE_TO_KEY = {
+    "faq": "faq_hit_count",
+    "rag": "rag_hit_count",
+    "llm": "llm_hit_count",
+    "fallback": "fallback_count",
+}
+
+
+class MetricsCollector:
+    def __init__(self, redis_client):
+        self.redis = redis_client
+
+    def _prefix(self, tenant_id: str) -> str:
+        return f"tenant:{tenant_id}:metrics"
+
+    def _p95_key(self, tenant_id: str) -> str:
+        return f"tenant:{tenant_id}:{P95_SORTED_SET}"
+
+    async def record(self, tenant_id: str, result: RoutingResult) -> None:
+        prefix = self._prefix(tenant_id)
+        p95_key = self._p95_key(tenant_id)
+
+        pipe = self.redis.pipeline()
+        pipe.hincrby(prefix, "total_count", 1)
+
+        source_key = _SOURCE_TO_KEY.get(result.source)
+        if source_key:
+            pipe.hincrby(prefix, source_key, 1)
+
+        if result.is_timeout:
+            pipe.hincrby(prefix, "timeout_count", 1)
+
+        pipe.hincrby(prefix, "response_ms_sum", result.elapsed_ms)
+
+        # p95 sorted set — score=elapsed_ms, member=unique id
+        import time
+        member = f"{time.time_ns()}"
+        pipe.zadd(p95_key, {member: result.elapsed_ms})
+        pipe.zremrangebyrank(p95_key, 0, -(P95_MAX_SIZE + 1))
+
+        await pipe.execute()
+
+    async def get_overview(self, tenant_id: str) -> dict:
+        prefix = self._prefix(tenant_id)
+        p95_key = self._p95_key(tenant_id)
+
+        raw = await self.redis.hgetall(prefix)
+        counts = {k: int(v) for k, v in raw.items()} if raw else {}
+
+        total = counts.get("total_count", 0)
+        avg_ms = counts.get("response_ms_sum", 0) // max(total, 1)
+
+        # p95 계산
+        p95_ms = 0
+        buf_size = await self.redis.zcard(p95_key)
+        if buf_size > 0:
+            p95_idx = max(0, int(buf_size * 0.95) - 1)
+            p95_items = await self.redis.zrange(p95_key, p95_idx, p95_idx, withscores=True)
+            if p95_items:
+                p95_ms = int(p95_items[0][1])
+
+        rates = {}
+        if total > 0:
+            rates["faq_hit_rate"] = round(counts.get("faq_hit_count", 0) / total * 100, 2)
+            rates["rag_hit_rate"] = round(counts.get("rag_hit_count", 0) / total * 100, 2)
+            rates["fallback_rate"] = round(counts.get("fallback_count", 0) / total * 100, 2)
+            rates["timeout_rate"] = round(counts.get("timeout_count", 0) / total * 100, 2)
+        else:
+            rates = {"faq_hit_rate": 0.0, "rag_hit_rate": 0.0, "fallback_rate": 0.0, "timeout_rate": 0.0}
+
+        return {
+            "counts": counts,
+            "rates": rates,
+            "avg_ms": avg_ms,
+            "p95_ms": p95_ms,
+        }

backend/app/services/moderation.py

@@ -0,0 +1,175 @@
+"""
+악성·반복 민원 제한 서비스.
+
+Level  상태         자동 조치              해제
+0      정상         없음                   자동
+1      주의         경고 메시지            자동
+2      경고         30초 응답 지연         자동 24h
+3      제한         10회/일 제한           자동 72h
+4      임시 차단    24시간 차단            편집장 수동 확인 필요
+5      영구 제한    차단 유지              편집장 수동 해제만
+
+원칙: 자동 영구 차단 없음. Level 4+ 는 편집장 수동 확인.
+"""
+import asyncio
+from datetime import datetime, timedelta, timezone
+from typing import Optional
+
+from sqlalchemy import select
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from app.models.moderation import UserRestriction, RestrictionLevel
+
+# 레벨별 자동 만료 시간
+LEVEL_EXPIRY = {
+    RestrictionLevel.WARNING: timedelta(hours=24),
+    RestrictionLevel.LIMITED: timedelta(hours=72),
+    RestrictionLevel.SUSPENDED: timedelta(hours=24),  # 편집장 확인 전 임시
+}
+
+# 레벨 3 일별 제한 횟수
+DAILY_LIMIT = 10
+
+# 레벨 1 경고 메시지
+WARNING_MESSAGE = "⚠️ 동일 문의가 반복되고 있습니다. 잠시 후 다시 시도해 주세요."
+
+
+class ModerationResult:
+    def __init__(
+        self,
+        allowed: bool,
+        level: int = 0,
+        message: Optional[str] = None,
+        delay_seconds: int = 0,
+    ):
+        self.allowed = allowed
+        self.level = level
+        self.message = message  # 경고 메시지 (Level 1)
+        self.delay_seconds = delay_seconds  # 응답 지연 (Level 2)
+
+
+class ModerationService:
+    def __init__(self, db: AsyncSession):
+        self.db = db
+
+    async def check(self, tenant_id: str, user_key: str) -> ModerationResult:
+        """
+        user_key의 제한 레벨 확인.
+        만료된 제한은 자동 해제.
+        """
+        restriction = await self._get_restriction(tenant_id, user_key)
+
+        if restriction is None:
+            return ModerationResult(allowed=True, level=0)
+
+        # 만료 확인
+        if restriction.expires_at and restriction.expires_at < datetime.now(timezone.utc):
+            if restriction.level < RestrictionLevel.SUSPENDED:
+                await self._reset(restriction)
+                return ModerationResult(allowed=True, level=0)
+
+        level = restriction.level
+
+        if level == RestrictionLevel.BLOCKED:
+            return ModerationResult(allowed=False, level=level)
+
+        if level == RestrictionLevel.SUSPENDED:
+            # Level 4: 편집장 확인 전 차단
+            return ModerationResult(allowed=False, level=level)
+
+        if level == RestrictionLevel.LIMITED:
+            # Level 3: 일별 10회 제한 (Redis 카운터 없이 단순 차단)
+            return ModerationResult(
+                allowed=False,
+                level=level,
+                message="일일 문의 한도에 도달했습니다. 내일 다시 시도해 주세요.",
+            )
+
+        if level == RestrictionLevel.WARNING:
+            # Level 2: 30초 지연
+            return ModerationResult(
+                allowed=True,
+                level=level,
+                delay_seconds=30,
+                message="요청이 지연되고 있습니다.",
+            )
+
+        if level == RestrictionLevel.NORMAL + 1:  # Level 1 (WARNING 사용하기 전 단계는 없으므로 1 = 주의)
+            return ModerationResult(
+                allowed=True,
+                level=level,
+                message=WARNING_MESSAGE,
+            )
+
+        return ModerationResult(allowed=True, level=level)
+
+    async def escalate(
+        self,
+        tenant_id: str,
+        user_key: str,
+        reason: str = "자동 감지",
+    ) -> int:
+        """
+        레벨 1단계 상승. Level 4+ 는 편집장 수동 확인.
+        반환: 새 레벨.
+        """
+        restriction = await self._get_restriction(tenant_id, user_key)
+
+        if restriction is None:
+            restriction = UserRestriction(
+                tenant_id=tenant_id,
+                user_key=user_key,
+                level=RestrictionLevel.NORMAL,
+                auto_applied=True,
+            )
+            self.db.add(restriction)
+
+        current = restriction.level
+        if current >= RestrictionLevel.SUSPENDED:
+            # Level 4+ 는 자동 상승 금지
+            return current
+
+        new_level = min(current + 1, RestrictionLevel.SUSPENDED)
+        restriction.level = new_level
+        restriction.reason = reason
+        restriction.auto_applied = True
+
+        # 만료 시간 설정
+        expiry_delta = LEVEL_EXPIRY.get(new_level)
+        if expiry_delta:
+            restriction.expires_at = datetime.now(timezone.utc) + expiry_delta
+        else:
+            restriction.expires_at = None
+
+        await self.db.commit()
+        return new_level
+
+    async def release(
+        self,
+        tenant_id: str,
+        user_key: str,
+        applied_by: str,
+    ) -> None:
+        """수동 해제 (편집장 이상)."""
+        restriction = await self._get_restriction(tenant_id, user_key)
+        if restriction:
+            await self._reset(restriction, applied_by=applied_by)
+
+    async def _get_restriction(
+        self, tenant_id: str, user_key: str
+    ) -> Optional[UserRestriction]:
+        result = await self.db.execute(
+            select(UserRestriction).where(
+                UserRestriction.tenant_id == tenant_id,
+                UserRestriction.user_key == user_key,
+            )
+        )
+        return result.scalar_one_or_none()
+
+    async def _reset(self, restriction: UserRestriction, applied_by: Optional[str] = None) -> None:
+        restriction.level = RestrictionLevel.NORMAL
+        restriction.expires_at = None
+        restriction.auto_applied = applied_by is None
+        if applied_by:
+            restriction.applied_by = applied_by
+        await self.db.commit()

backend/app/services/parsers/text_parser.py

@@ -0,0 +1,96 @@
+"""
+문서 파서 — 1차 정식 지원 형식:
+TXT · MD · DOCX · 텍스트 PDF · HTML
+"""
+import io
+from typing import Optional
+
+
+def parse_txt(content: bytes, encoding: str = "utf-8") -> str:
+    return content.decode(encoding, errors="replace")
+
+
+def parse_md(content: bytes) -> str:
+    return content.decode("utf-8", errors="replace")
+
+
+def parse_html(content: bytes) -> str:
+    from bs4 import BeautifulSoup
+    soup = BeautifulSoup(content, "html.parser")
+    # script/style 제거
+    for tag in soup(["script", "style"]):
+        tag.decompose()
+    return soup.get_text(separator="\n", strip=True)
+
+
+def parse_docx(content: bytes) -> str:
+    from docx import Document
+    doc = Document(io.BytesIO(content))
+    return "\n".join(p.text for p in doc.paragraphs if p.text.strip())
+
+
+def parse_pdf(content: bytes) -> str:
+    try:
+        import pdfplumber
+        with pdfplumber.open(io.BytesIO(content)) as pdf:
+            pages = []
+            for page in pdf.pages:
+                text = page.extract_text()
+                if text:
+                    pages.append(text)
+            return "\n".join(pages)
+    except Exception:
+        return ""
+
+
+PARSERS = {
+    "txt": parse_txt,
+    "md": parse_md,
+    "html": parse_html,
+    "htm": parse_html,
+    "docx": parse_docx,
+    "pdf": parse_pdf,
+}
+
+
+def extract_text(content: bytes, filename: str) -> Optional[str]:
+    """파일 확장자에 따라 적절한 파서 선택."""
+    ext = filename.rsplit(".", 1)[-1].lower() if "." in filename else ""
+    parser = PARSERS.get(ext)
+    if not parser:
+        return None
+    try:
+        return parser(content)
+    except Exception:
+        return None
+
+
+def chunk_text(text: str, chunk_size: int = 500, overlap: int = 50) -> list[str]:
+    """
+    문단 단위 청킹 (약 chunk_size 토큰).
+    overlap: 이전 청크 끝 글자를 다음 청크 시작에 포함.
+    """
+    paragraphs = [p.strip() for p in text.split("\n") if p.strip()]
+    chunks = []
+    current = []
+    current_len = 0
+
+    for para in paragraphs:
+        para_len = len(para)
+        if current_len + para_len > chunk_size and current:
+            chunk_text_ = "\n".join(current)
+            chunks.append(chunk_text_)
+            # overlap: 마지막 문단 유지
+            if overlap > 0 and current:
+                current = [current[-1]]
+                current_len = len(current[-1])
+            else:
+                current = []
+                current_len = 0
+        current.append(para)
+        current_len += para_len
+
+    if current:
+        chunks.append("\n".join(current))
+
+    return chunks if chunks else [text[:chunk_size]]

backend/app/services/rag_search.py

@@ -0,0 +1,109 @@
+"""
+Tier B — RAG 검색.
+임베딩 유사도 ≥ 0.70 + 근거 문서 존재 → 문서 기반 템플릿 응답.
+"""
+from typing import Optional
+
+from sqlalchemy import select
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from app.models.knowledge import Document
+from app.providers.base import SearchResult
+from app.providers.embedding import EmbeddingProvider
+from app.providers.vectordb import VectorDBProvider
+
+RAG_SIMILARITY_THRESHOLD = 0.70  # Tier B 기준
+
+
+class RAGSearchResult:
+    def __init__(self, chunk_text: str, doc: Document, score: float):
+        self.chunk_text = chunk_text
+        self.doc = doc
+        self.score = score
+
+    @property
+    def doc_name(self) -> str:
+        return self.doc.filename
+
+    @property
+    def doc_date(self) -> str:
+        if self.doc.published_at:
+            return self.doc.published_at.strftime("%Y.%m")
+        return ""
+
+
+class RAGSearchService:
+    def __init__(
+        self,
+        embedding_provider: EmbeddingProvider,
+        vectordb_provider: VectorDBProvider,
+        db: AsyncSession,
+    ):
+        self.embedding = embedding_provider
+        self.vectordb = vectordb_provider
+        self.db = db
+
+    async def search(
+        self, tenant_id: str, utterance: str, top_k: int = 3
+    ) -> Optional[list[RAGSearchResult]]:
+        """
+        발화를 임베딩 → 벡터DB 검색 → 0.70 이상 문서 청크 반환.
+        결과 없으면 None.
+        """
+        try:
+            vecs = await self.embedding.embed([utterance])
+        except NotImplementedError:
+            return None
+
+        query_vec = vecs[0]
+        results = await self.vectordb.search(
+            tenant_id=tenant_id,
+            query_vec=query_vec,
+            top_k=top_k,
+            threshold=RAG_SIMILARITY_THRESHOLD,
+        )
+
+        if not results:
+            return None
+
+        # 중복 doc_id 제거 (같은 문서의 여러 청크 중 최고 점수만)
+        seen_docs: dict[str, SearchResult] = {}
+        for r in results:
+            doc_id = r.metadata.get("doc_id", r.doc_id.rsplit("_", 1)[0])
+            if doc_id not in seen_docs or r.score > seen_docs[doc_id].score:
+                seen_docs[doc_id] = r
+
+        # Document 레코드 로드 (is_active=True만)
+        rag_results = []
+        for doc_id, sr in seen_docs.items():
+            doc = await self._load_doc(tenant_id, doc_id)
+            if doc:
+                rag_results.append(RAGSearchResult(sr.text, doc, sr.score))
+
+        return rag_results if rag_results else None
+
+    async def _load_doc(self, tenant_id: str, doc_id: str) -> Optional[Document]:
+        result = await self.db.execute(
+            select(Document).where(
+                Document.tenant_id == tenant_id,
+                Document.id == doc_id,
+                Document.is_active.is_(True),
+            )
+        )
+        return result.scalar_one_or_none()
+
+    def build_answer(self, utterance: str, rag_results: list[RAGSearchResult]) -> str:
+        """
+        문서 기반 템플릿 응답 생성.
+        출처 2단계 포맷 (간단형).
+        """
+        # 근거 문단 합치기 (최대 2개)
+        contexts = [r.chunk_text[:300] for r in rag_results[:2]]
+        context_str = "\n---\n".join(contexts)
+
+        best = rag_results[0]
+        citation = f"📎 출처: {best.doc_name}"
+        if best.doc_date:
+            citation += f" ({best.doc_date})"
+
+        return f"{context_str}\n\n{citation}"

backend/app/services/routing.py

@@ -0,0 +1,243 @@
+import asyncio
+from dataclasses import dataclass, field
+from typing import Optional
+
+
+@dataclass
+class RoutingResult:
+    answer: str
+    tier: str  # 'A'|'B'|'C'|'D'
+    source: str  # 'faq'|'rag'|'llm'|'fallback'
+    faq_id: Optional[str] = None
+    doc_id: Optional[str] = None
+    doc_name: Optional[str] = None
+    doc_date: Optional[str] = None
+    score: float = 0.0
+    elapsed_ms: int = 0
+    is_timeout: bool = False
+    request_id: Optional[str] = None
+
+    def to_dict(self) -> dict:
+        citations = []
+        if self.doc_name:
+            citations.append({"doc": self.doc_name, "date": self.doc_date or ""})
+        return {
+            "answer": self.answer,
+            "tier": self.tier,
+            "source": self.source,
+            "faq_id": self.faq_id,
+            "doc_id": self.doc_id,
+            "score": self.score,
+            "elapsed_ms": self.elapsed_ms,
+            "is_timeout": self.is_timeout,
+            "request_id": self.request_id,
+            "citations": citations,
+        }
+
+
+class ResponseRouter:
+    TIMEOUT_MS = 4500  # 4.5초 — 카카오 5초 한계 - 500ms
+
+    def __init__(self, tenant_config: dict, providers: dict):
+        self.tenant_config = tenant_config
+        self.providers = providers
+
+    async def route(
+        self,
+        tenant_id: str,
+        utterance: str,
+        user_key: str,
+        request_id: Optional[str] = None,
+        db=None,
+    ) -> RoutingResult:
+        import time
+        start = time.monotonic()
+        try:
+            return await asyncio.wait_for(
+                self._try_tiers(tenant_id, utterance, user_key, request_id, db, start),
+                timeout=self.TIMEOUT_MS / 1000,
+            )
+        except asyncio.TimeoutError:
+            elapsed = int((time.monotonic() - start) * 1000)
+            return self._tier_d(tenant_id, elapsed, is_timeout=True, request_id=request_id)
+
+    async def _try_tiers(
+        self,
+        tenant_id: str,
+        utterance: str,
+        user_key: str,
+        request_id: Optional[str],
+        db,
+        start: float,
+    ) -> RoutingResult:
+        import time
+
+        # Tier A — FAQ 임베딩 유사도 검색
+        tier_a = await self._try_tier_a(tenant_id, utterance, db)
+        if tier_a is not None:
+            tier_a.elapsed_ms = int((time.monotonic() - start) * 1000)
+            tier_a.request_id = request_id
+            return tier_a
+
+        # Tier C — LLM 기반 재서술 (RAG 근거 있음 + LLM 활성화)
+        # Tier B보다 먼저 시도: LLM 활성 시 템플릿 대신 재서술
+        tier_c = await self._try_tier_c(tenant_id, utterance, db)
+        if tier_c is not None:
+            tier_c.elapsed_ms = int((time.monotonic() - start) * 1000)
+            tier_c.request_id = request_id
+            return tier_c
+
+        # Tier B — RAG 문서 검색 (LLM 비활성 또는 Tier C 실패 시)
+        tier_b = await self._try_tier_b(tenant_id, utterance, db)
+        if tier_b is not None:
+            tier_b.elapsed_ms = int((time.monotonic() - start) * 1000)
+            tier_b.request_id = request_id
+            return tier_b
+
+        elapsed = int((time.monotonic() - start) * 1000)
+        return self._tier_d(tenant_id, elapsed, request_id=request_id)
+
+    async def _try_tier_a(self, tenant_id: str, utterance: str, db) -> Optional[RoutingResult]:
+        """Tier A — FAQ 임베딩 유사도 ≥ 0.85."""
+        embedding_provider = self.providers.get("embedding")
+        vectordb_provider = self.providers.get("vectordb")
+
+        if embedding_provider is None or vectordb_provider is None or db is None:
+            return None
+
+        from app.services.faq_search import FAQSearchService
+        service = FAQSearchService(embedding_provider, vectordb_provider, db)
+        match = await service.search(tenant_id, utterance)
+
+        if match is None:
+            return None
+
+        faq, score = match
+        # hit_count 비동기 증가 (fire-and-forget)
+        await service.increment_hit(faq.id)
+
+        citation_date = (
+            faq.updated_at.strftime("%Y.%m") if faq.updated_at else ""
+        )
+
+        return RoutingResult(
+            answer=faq.answer,
+            tier="A",
+            source="faq",
+            faq_id=faq.id,
+            doc_name=f"FAQ: {faq.question[:30]}",
+            doc_date=citation_date,
+            score=score,
+        )
+
+    async def _try_tier_c(self, tenant_id: str, utterance: str, db) -> Optional[RoutingResult]:
+        """Tier C — RAG 근거 있음 + LLM 활성화 → 근거 기반 재서술."""
+        llm_provider = self.providers.get("llm")
+        embedding_provider = self.providers.get("embedding")
+        vectordb_provider = self.providers.get("vectordb")
+
+        if llm_provider is None or embedding_provider is None or vectordb_provider is None or db is None:
+            return None
+
+        # NullLLMProvider → None 즉시 반환
+        from app.providers.llm import NullLLMProvider
+        if isinstance(llm_provider, NullLLMProvider):
+            return None
+
+        # RAG 검색 (Tier B와 동일 임계값)
+        from app.services.rag_search import RAGSearchService
+        rag_service = RAGSearchService(embedding_provider, vectordb_provider, db)
+        rag_results = await rag_service.search(tenant_id, utterance)
+
+        if not rag_results:
+            return None  # 근거 없으면 LLM 미호출 (P6 할루시네이션 방지)
+
+        # 근거 기반 LLM 재서술
+        context_chunks = [r.chunk_text for r in rag_results[:3]]
+        context_str = "\n---\n".join(context_chunks)
+        tenant_name = self.tenant_config.get("tenant_name", "")
+        name_prefix = f"{tenant_name}의 " if tenant_name else ""
+
+        system_prompt = (
+            f"당신은 {name_prefix}AI 안내 도우미입니다.\n"
+            f"반드시 아래 근거 문서의 내용만을 바탕으로 답변하세요.\n"
+            f"근거 없는 내용은 절대 추측하지 마세요.\n\n"
+            f"근거 문서:\n{context_str}"
+        )
+
+        answer = await llm_provider.generate(
+            system_prompt=system_prompt,
+            user_message=utterance,
+            context_chunks=context_chunks,
+        )
+
+        if answer is None:
+            return None  # LLM 실패 → Tier D로 폴백
+
+        best = rag_results[0]
+        return RoutingResult(
+            answer=answer,
+            tier="C",
+            source="llm",
+            doc_id=best.doc.id,
+            doc_name=best.doc_name,
+            doc_date=best.doc_date,
+            score=best.score,
+        )
+
+    async def _try_tier_b(self, tenant_id: str, utterance: str, db) -> Optional[RoutingResult]:
+        """Tier B — RAG 유사도 ≥ 0.70 + 근거 문서 존재."""
+        embedding_provider = self.providers.get("embedding")
+        vectordb_provider = self.providers.get("vectordb")
+
+        if embedding_provider is None or vectordb_provider is None or db is None:
+            return None
+
+        from app.services.rag_search import RAGSearchService
+        service = RAGSearchService(embedding_provider, vectordb_provider, db)
+        results = await service.search(tenant_id, utterance)
+
+        if not results:
+            return None
+
+        best = results[0]
+        answer = service.build_answer(utterance, results)
+
+        return RoutingResult(
+            answer=answer,
+            tier="B",
+            source="rag",
+            doc_id=best.doc.id,
+            doc_name=best.doc_name,
+            doc_date=best.doc_date,
+            score=best.score,
+        )
+
+    def _tier_d(
+        self,
+        tenant_id: str,
+        elapsed_ms: int,
+        is_timeout: bool = False,
+        request_id: Optional[str] = None,
+    ) -> RoutingResult:
+        # DB 조회 없이 tenant_config 메모리에서 직접 읽음 (~5ms)
+        phone = self.tenant_config.get("phone_number", "")
+        contact = self.tenant_config.get("fallback_dept", "")
+        name = self.tenant_config.get("tenant_name", "")
+
+        if phone and contact:
+            answer = f"해당 문의는 {name} {contact}({phone})로 연락해 주세요."
+        elif phone:
+            answer = f"해당 문의는 {name}({phone})로 연락해 주세요." if name else f"해당 문의는 {phone}로 연락해 주세요."
+        elif name:
+            answer = f"죄송합니다. {name}에 직접 문의해 주세요."
+        else:
+            answer = "죄송합니다. 해당 내용을 찾을 수 없습니다. 담당자에게 직접 문의해 주세요."
+        return RoutingResult(
+            answer=answer,
+            tier="D",
+            source="fallback",
+            elapsed_ms=elapsed_ms,
+            is_timeout=is_timeout,
+            request_id=request_id,
+        )