perf(interview): 꼬리질문 첫 토큰 지연 단축 + 타자기 연출

csh1668 · csh1668 · commit 669fd1141949 · 2026-06-05T00:24:03.000+09:00
P1(AI): followup RAG에서 실패·저속한 LLM 리랭커를 기본 비활성화(followup_rerank_enabled=false)하고 candidate 확장 없이 top_k 직접 검색. _build_rag_context 전체에 타임아웃(followup_rag_timeout_sec=1.5s) 폴백 추가 → 첫 토큰까지 ~3s에서 ~1.5s 목표.

P2(FE): SESSION_MESSAGE_DELTA가 짧은 질문에서 0.3s에 몰려 와도 보이도록 useTypewriter로 누적 텍스트를 점진 표시(StageQuestion/QuestionBubble). 긴 스트림은 남은 양 비례 step으로 따라잡아 지연 없음.
diff --git a/ai/src/ai_server/config/settings.py b/ai/src/ai_server/config/settings.py
@@ -41,6 +41,10 @@ class Settings(BaseSettings):
     # 리랭킹: 하이브리드 검색으로 후보 N개(candidate_k)를 가져와 LLM 으로 재정렬 후 top_k 주입.
     rerank_enabled: bool = True
     rerank_candidate_k: int = 20
+    # 꼬리질문 RAG 저지연: 리랭커(LLM 호출, 종종 실패)를 건너뛰고 검색 순서를 그대로 쓴다.
+    followup_rerank_enabled: bool = False
+    # RAG 컨텍스트 구성(임베딩+검색+리랭크) 전체 상한. 초과 시 (none) 으로 폴백해 첫 토큰을 막지 않는다.
+    followup_rag_timeout_sec: float = 1.5
     # 질문 풀 초기 크기. Core 의 applyPool 이 questions[0] 만 INSERT 하므로 1 로 고정해 토큰 낭비 차단.
     # 후속 작업에서 풀 저장 도입 시 늘리기 (예: 5).
     questions_initial_pool_size: int = 1
diff --git a/ai/src/ai_server/messaging/consumers/followup_consumer.py b/ai/src/ai_server/messaging/consumers/followup_consumer.py
@@ -50,6 +50,8 @@ def __init__(
         tts=None,
         storage=None,
         tts_voice: str = "",
+        rerank_enabled: bool = True,
+        rag_timeout_sec: float = 1.5,
     ) -> None:
         self._generator = generator
         self._publisher = publisher
@@ -65,6 +67,8 @@ def __init__(
         self._tts = tts
         self._storage = storage
         self._tts_voice = tts_voice
+        self._rerank_enabled = rerank_enabled
+        self._rag_timeout_sec = rag_timeout_sec
 
     async def handle(self, message: AbstractIncomingMessage) -> None:
         async with message.process(requeue=False):
@@ -233,8 +237,22 @@ async def _synth_segment(
     async def _build_rag_context(self, req: GenerateFollowupRequest) -> str:
         if not self._core or not self._embedder or not req.context_document_ids:
             return "(none)"
+        try:
+            return await asyncio.wait_for(
+                self._do_build_rag_context(req), timeout=self._rag_timeout_sec
+            )
+        except asyncio.TimeoutError:
+            log.warning(
+                "followup.rag.timeout",
+                session_id=req.session_id,
+                timeout_sec=self._rag_timeout_sec,
+            )
+            return "(none)"
 
+    async def _do_build_rag_context(self, req: GenerateFollowupRequest) -> str:
         query = f"{req.previous_question}\n\n{req.answer_text}"
+        # 리랭크 비활성 시 candidate 확장 없이 top_k 만 검색(LLM 리랭커 호출 회피).
+        search_k = self._candidate_k if self._rerank_enabled else self._rag_top_k
         try:
             query_vec = (
                 await self._embedder.embed([query], task_type="RETRIEVAL_QUERY")
@@ -243,17 +261,18 @@ async def _build_rag_context(self, req: GenerateFollowupRequest) -> str:
                 query_embedding=query_vec,
                 query_text=query,
                 document_ids=req.context_document_ids,
-                top_k=self._candidate_k,
+                top_k=search_k,
             )
         except Exception as exc:
             log.warn("followup.rag.failed", error=str(exc), session_id=req.session_id)
             return "(none)"
 
         if not hits:
             return "(none)"
-        hits = await rerank_hits(
-            self._reranker, query=query, hits=hits, top_k=self._rag_top_k
-        )
+        if self._rerank_enabled:
+            hits = await rerank_hits(
+                self._reranker, query=query, hits=hits, top_k=self._rag_top_k
+            )
         return "\n---\n".join(
             f"[doc#{h.document_id} chunk#{h.chunk_index}] {h.chunk_text}" for h in hits
         )
diff --git a/ai/src/ai_server/messaging/runner.py b/ai/src/ai_server/messaging/runner.py
@@ -213,6 +213,8 @@ def __init__(self, settings: Settings) -> None:
             tts=tts,
             storage=storage,
             tts_voice=settings.openai_tts_voice,
+            rerank_enabled=settings.followup_rerank_enabled,
+            rag_timeout_sec=settings.followup_rag_timeout_sec,
         )
 
         # 종합 피드백 생성 (US-24)
diff --git a/ai/tests/test_followup_consumer.py b/ai/tests/test_followup_consumer.py
@@ -480,3 +480,163 @@ async def test_callback_includes_answer_message_id():
     payload = publisher.publish.await_args.kwargs["payload"]
     assert payload.answer_message_id == 502  # _envelope 의 answerMessageId
     assert payload.followup_message_id == 503  # _envelope 의 followupMessageId
+
+
+# ---------------------------------------------------------------------------
+# 꼬리질문 RAG 저지연 픽스: rerank_enabled=False, timeout 폴백
+# ---------------------------------------------------------------------------
+
+
+def _make_req() -> "GenerateFollowupRequest":
+    from ai_server.model.messages.followup import GenerateFollowupRequest
+
+    return GenerateFollowupRequest(
+        session_id=99,
+        parent_message_id=501,
+        answer_message_id=502,
+        followup_message_id=503,
+        previous_question="결제 outbox 어떻게 구현?",
+        answer_text="RabbitMQ로 보냈습니다.",
+        mode="TECHNICAL",
+        job_category="BACKEND",
+        context_document_ids=[1],
+    )
+
+
+def _make_hit(document_id: int = 1, chunk_index: int = 0, chunk_text: str = "x"):
+    from types import SimpleNamespace
+
+    return SimpleNamespace(
+        document_id=document_id, chunk_index=chunk_index, chunk_text=chunk_text
+    )
+
+
+@pytest.mark.asyncio
+async def test_rag_rerank_skipped_when_disabled():
+    """rerank_enabled=False: search_embeddings 는 top_k 만 요청, reranker.rerank 는 호출 안 됨."""
+    import asyncio
+
+    from ai_server.messaging.consumers.followup_consumer import FollowupConsumer
+    from ai_server.messaging.idempotency import LruIdempotencyStore
+
+    hit = _make_hit(chunk_text="이 청크가 반환돼야 한다")
+
+    embedder = MagicMock()
+    embedder.embed = AsyncMock(return_value=[[0.0]])
+
+    core = MagicMock()
+    core.search_embeddings = AsyncMock(return_value=[hit])
+
+    fake_reranker = MagicMock()
+    fake_reranker.rerank = AsyncMock(return_value=[0])
+
+    generator = MagicMock()
+    generator.generate = AsyncMock()
+    publisher = MagicMock()
+    publisher.publish = AsyncMock()
+
+    consumer = FollowupConsumer(
+        generator=generator,
+        publisher=publisher,
+        idempotency=LruIdempotencyStore(max_size=10),
+        callback_routing_key="callback.questions",
+        core_client=core,
+        embedder=embedder,
+        rag_top_k=5,
+        reranker=fake_reranker,
+        candidate_k=20,
+        rerank_enabled=False,
+    )
+
+    req = _make_req()
+    result = await consumer._build_rag_context(req)
+
+    # 청크 텍스트가 결과에 포함돼야 한다
+    assert "이 청크가 반환돼야 한다" in result
+    # rerank 비활성 시 candidate 확장 없이 rag_top_k 만 검색
+    call_kwargs = core.search_embeddings.await_args.kwargs
+    assert call_kwargs["top_k"] == 5  # rag_top_k, NOT candidate_k(20)
+    # 리랭커 rerank 는 절대 호출되지 않음
+    fake_reranker.rerank.assert_not_awaited()
+
+
+@pytest.mark.asyncio
+async def test_rag_rerank_used_when_enabled():
+    """rerank_enabled=True: search_embeddings 는 candidate_k, reranker.rerank 는 호출됨."""
+    hit = _make_hit(chunk_text="후보 청크")
+
+    embedder = MagicMock()
+    embedder.embed = AsyncMock(return_value=[[0.0]])
+
+    core = MagicMock()
+    core.search_embeddings = AsyncMock(return_value=[hit])
+
+    fake_reranker = MagicMock()
+    fake_reranker.rerank = AsyncMock(return_value=[0])
+
+    generator = MagicMock()
+    generator.generate = AsyncMock()
+    publisher = MagicMock()
+    publisher.publish = AsyncMock()
+
+    consumer = FollowupConsumer(
+        generator=generator,
+        publisher=publisher,
+        idempotency=LruIdempotencyStore(max_size=10),
+        callback_routing_key="callback.questions",
+        core_client=core,
+        embedder=embedder,
+        rag_top_k=5,
+        reranker=fake_reranker,
+        candidate_k=20,
+        rerank_enabled=True,
+    )
+
+    req = _make_req()
+    result = await consumer._build_rag_context(req)
+
+    assert "후보 청크" in result
+    call_kwargs = core.search_embeddings.await_args.kwargs
+    assert call_kwargs["top_k"] == 20  # candidate_k
+    # 리랭커 rerank 는 호출돼야 한다
+    fake_reranker.rerank.assert_awaited_once()
+
+
+@pytest.mark.asyncio
+async def test_rag_timeout_returns_none():
+    """rag_timeout_sec 초과 시 _build_rag_context 가 '(none)' 을 반환한다."""
+    import asyncio
+
+    from ai_server.messaging.consumers.followup_consumer import FollowupConsumer
+    from ai_server.messaging.idempotency import LruIdempotencyStore
+
+    async def _slow_embed(texts, *, task_type=""):
+        await asyncio.sleep(0.1)  # 타임아웃(0.01s) 보다 훨씬 길다
+        return [[0.0]]
+
+    embedder = MagicMock()
+    embedder.embed = _slow_embed
+
+    core = MagicMock()
+    core.search_embeddings = AsyncMock(return_value=[])
+
+    generator = MagicMock()
+    generator.generate = AsyncMock()
+    publisher = MagicMock()
+    publisher.publish = AsyncMock()
+
+    consumer = FollowupConsumer(
+        generator=generator,
+        publisher=publisher,
+        idempotency=LruIdempotencyStore(max_size=10),
+        callback_routing_key="callback.questions",
+        core_client=core,
+        embedder=embedder,
+        rag_top_k=5,
+        rag_timeout_sec=0.01,
+    )
+
+    req = _make_req()
+    result = await consumer._build_rag_context(req)
+
+    assert result == "(none)"
diff --git a/frontend/src/features/interview/lib/useTypewriter.test.ts b/frontend/src/features/interview/lib/useTypewriter.test.ts
@@ -0,0 +1,22 @@
+import { describe, it, expect, vi, beforeEach, afterEach } from 'vitest'
+import { renderHook, act } from '@testing-library/react'
+import { useTypewriter } from './useTypewriter'
+
+describe('useTypewriter', () => {
+  beforeEach(() => vi.useFakeTimers())
+  afterEach(() => vi.useRealTimers())
+
+  it('enabled=false 면 전체 텍스트 즉시', () => {
+    const { result } = renderHook(() => useTypewriter('안녕하세요', false))
+    expect(result.current).toBe('안녕하세요')
+  })
+
+  it('enabled 면 점진적으로 드러나 결국 전체에 도달', () => {
+    const { result } = renderHook(() => useTypewriter('가나다라마바사아자차', true))
+    // 처음엔 일부만(또는 비어있음)
+    const initial = result.current.length
+    act(() => { vi.advanceTimersByTime(35 * 12) })
+    expect(result.current).toBe('가나다라마바사아자차')
+    expect(initial).toBeLessThan('가나다라마바사아자차'.length)
+  })
+})
diff --git a/frontend/src/features/interview/lib/useTypewriter.ts b/frontend/src/features/interview/lib/useTypewriter.ts
@@ -0,0 +1,29 @@
+import { useEffect, useRef, useState } from 'react'
+
+// 스트리밍 텍스트를 일정 속도로 점진 표시(타자기). enabled=false면 전체 즉시 반환.
+// fullText 가 늘어나면 따라가고, 남은 양에 비례해 step 을 키워 길고 빠른 스트림은 지연 없이 따라잡는다.
+export function useTypewriter(fullText: string, enabled: boolean): string {
+  const [count, setCount] = useState(enabled ? 0 : fullText.length)
+  const targetRef = useRef(fullText)
+  targetRef.current = fullText
+
+  useEffect(() => {
+    if (!enabled) {
+      setCount(targetRef.current.length)
+      return
+    }
+    const id = setInterval(() => {
+      setCount((c) => {
+        const target = targetRef.current.length
+        if (c >= target) return c
+        const remaining = target - c
+        const step = Math.max(2, Math.ceil(remaining / 6))
+        return Math.min(target, c + step)
+      })
+    }, 35)
+    return () => clearInterval(id)
+  }, [enabled])
+
+  const shown = enabled ? Math.min(count, fullText.length) : fullText.length
+  return fullText.slice(0, shown)
+}
diff --git a/frontend/src/features/interview/model/useLiveInterview.ts b/frontend/src/features/interview/model/useLiveInterview.ts
@@ -15,7 +15,7 @@ import type { OptimisticAnswer } from './optimistic'
 import { applyDelta, isStreamingMessage, FOLLOWUP_GENERATING_TEXT } from './streamingBuffer'
 import type { DeltaPayload } from './streamingBuffer'
 
-export type ThreadItem = Message & { key: string }
+export type ThreadItem = Message & { key: string; streaming?: boolean }
 export type ConnectionStatus = 'connecting' | 'open' | 'closed'
 
 export function useLiveInterview(sessionId: number) {
@@ -58,7 +58,7 @@ export function useLiveInterview(sessionId: number) {
   const mergedMessages = serverMessages.map((m) => {
     const buffered = deltaBuffer[m.id ?? -1]
     if (buffered !== undefined && isStreamingMessage(m, buffered)) {
-      return { ...m, content: buffered }
+      return { ...m, content: buffered, streaming: true as const }
     }
     return m
   })
diff --git a/frontend/src/features/interview/ui/live/ConversationThread.tsx b/frontend/src/features/interview/ui/live/ConversationThread.tsx
@@ -27,7 +27,7 @@ export function ConversationThread({
     <div ref={containerRef} className="flex h-full flex-col gap-3 overflow-y-auto px-4 py-6">
       {items.map((item) =>
         isQuestion(item) ? (
-          <QuestionBubble key={item.key} message={item} autoPlay={item.key === lastQuestionKey} />
+          <QuestionBubble key={item.key} message={item} autoPlay={item.key === lastQuestionKey} streaming={item.streaming} />
         ) : (
           <AnswerBubble key={item.key} message={item} />
         ),
diff --git a/frontend/src/features/interview/ui/live/InterviewStage.tsx b/frontend/src/features/interview/ui/live/InterviewStage.tsx
@@ -119,7 +119,7 @@ export function InterviewStage({
         {awaitingQuestion || !currentQuestion ? (
           <ThinkingState transcribing={transcribing} />
         ) : (
-          <StageQuestion question={currentQuestion} segmented={wasSegmented(currentQuestion.id ?? -1)} />
+          <StageQuestion question={currentQuestion} segmented={wasSegmented(currentQuestion.id ?? -1)} streaming={currentQuestion?.streaming ?? false} />
         )}
       </div>
 
diff --git a/frontend/src/features/interview/ui/live/QuestionBubble.tsx b/frontend/src/features/interview/ui/live/QuestionBubble.tsx
@@ -3,6 +3,7 @@ import { StatusBadge } from '@/shared/ui/StatusBadge'
 import { categoryLabel } from '../../lib/categoryLabel'
 import { useTtsPlayback } from '../../lib/media/useTtsPlayback'
 import { FOLLOWUP_GENERATING_TEXT } from '../../model/streamingBuffer'
+import { useTypewriter } from '../../lib/useTypewriter'
 
 function PlayIcon({ playing }: { playing: boolean }) {
   return (
@@ -15,13 +16,17 @@ function PlayIcon({ playing }: { playing: boolean }) {
 export function QuestionBubble({
   message,
   autoPlay = false,
+  streaming = false,
 }: {
   message: Message
   autoPlay?: boolean
+  streaming?: boolean
 }) {
   const label = categoryLabel(message.category)
   const hasMeta = Boolean(label || message.targetEvidence)
   const ttsReady = message.ttsStatus === 'SUCCEEDED'
+  const isSentinel = message.content === FOLLOWUP_GENERATING_TEXT
+  const shownText = useTypewriter(message.content ?? '', !!streaming && !isSentinel)
 
   const { playing, toggle, audioNode } = useTtsPlayback({
     sessionId: message.sessionId,
@@ -42,14 +47,14 @@ export function QuestionBubble({
           </div>
         )}
         <div className="rounded-lg rounded-tl-sm bg-surface-raised px-4 py-3 text-body text-fg shadow-sm">
-          {message.content === FOLLOWUP_GENERATING_TEXT ? (
+          {isSentinel ? (
             <span className="inline-flex gap-1 text-fg-muted" aria-label="질문 생성 중">
               <span className="animate-pulse">●</span>
               <span className="animate-pulse [animation-delay:150ms]">●</span>
               <span className="animate-pulse [animation-delay:300ms]">●</span>
             </span>
           ) : (
-            <p className="whitespace-pre-wrap">{message.content}</p>
+            <p className="whitespace-pre-wrap">{shownText}</p>
           )}
           {ttsReady && (
             <div className="mt-2 flex items-center gap-2 border-t border-border pt-2">
diff --git a/frontend/src/features/interview/ui/live/StageQuestion.tsx b/frontend/src/features/interview/ui/live/StageQuestion.tsx

Original file line number	Diff line number	Diff line change
`@@ -213,6 +213,8 @@ def __init__(self, settings: Settings) -> None:`
`213`	`213`	`tts=tts,`
`214`	`214`	`storage=storage,`
`215`	`215`	`tts_voice=settings.openai_tts_voice,`
	`216`	`+ rerank_enabled=settings.followup_rerank_enabled,`
	`217`	`+ rag_timeout_sec=settings.followup_rag_timeout_sec,`
`216`	`218`	`)`
`217`	`219`
`218`	`220`	`# 종합 피드백 생성 (US-24)`