feat(question): 초기 질문에 근거(target_evidence)·평가시그널 추가 + PLAN/Retrieve 분기

ai/src/ai_server/chain/prompts/question_generation.py

@@ -28,7 +28,13 @@
     "   - TECHNICAL: CS_FUNDAMENTAL + TECH_CHOICE + PROJECT_DEEP_DIVE 중심.\n"
     "   - PERSONALITY: BEHAVIORAL 중심.\n"
     "   - INTEGRATED: 모두 균형.\n"
-    "4. 지원자 컨텍스트에 명시된 기술/프로젝트를 구체적으로 언급하는 PROJECT_DEEP_DIVE / "
-    "TECH_CHOICE 질문을 최소 절반 이상 포함하세요.\n\n"
+    "   자료가 받쳐주는 만큼만 분배하고, 근거 없는 카테고리를 억지로 채우지 마세요.\n"
+    "4. 각 질문에 target_evidence 와 expected_signal 을 채웁니다:\n"
+    "   - target_evidence: 질문의 근거가 된 컨텍스트 원문을 짧게 인용. "
+    "PROJECT_DEEP_DIVE·TECH_CHOICE 는 **필수**(근거를 못 찾으면 그 질문을 만들지 마세요). "
+    "CS_FUNDAMENTAL·BEHAVIORAL 은 선택(없으면 빈 문자열).\n"
+    "   - expected_signal: 좋은 답변이 드러내야 할 핵심(평가 관점). 한 문장.\n"
+    "5. 지원자 컨텍스트의 기술/프로젝트를 구체적으로 언급하는 PROJECT_DEEP_DIVE / "
+    "TECH_CHOICE 질문을 우선합니다.\n\n"
     "{format_instructions}"
 )

ai/src/ai_server/messaging/consumers/questions_consumer.py

@@ -123,6 +123,11 @@ async def _build_context(self, req: GenerateQuestionsRequest) -> str:
         if not document_ids:
             return base_context
 
+        # PLAN vs Retrieve: 문서가 1개면 markdown 전체가 이미 base_context 에 들어 있어
+        # 검색이 중복이다 → 검색 생략(PLAN). 여러 문서일 때만 RAG 로 관련 청크를 추린다.
+        if len(document_ids) <= 1:
+            return base_context
+
         query = _build_initial_rag_query(req)
         try:
             query_vec = (

ai/src/ai_server/model/messages/questions.py

@@ -17,6 +17,7 @@
 
 class DocumentContext(BaseModel):
     """Core 가 generate.questions envelope 에 담아 보내는 문서 컨텍스트."""
+
     model_config = camel_config()
 
     document_id: int
@@ -39,10 +40,15 @@ class GenerateQuestionsRequest(BaseModel):
 
 class GeneratedQuestion(BaseModel):
     """LLM 응답 단위. category 는 Spring 측 enum 과 동기."""
+
     model_config = camel_config()
 
     category: QuestionCategory
     question: str
+    # 질문이 근거한 자료 인용(PROJECT/TECH 는 필수). 라이브 화면에 힌트로 노출.
+    target_evidence: str = ""
+    # 좋은 답이 드러내야 할 것 — 내부 평가용. 라이브 비노출(정답 유출 방지).
+    expected_signal: str = ""
 
 
 class QuestionPoolCallbackPayload(BaseModel):

ai/tests/test_questions_consumer.py

@@ -56,8 +56,12 @@ async def test_consumer_generates_questions_and_publishes_callback():
     generator.generate = AsyncMock(
         return_value=GeneratedQuestionPool(
             questions=[
-                GeneratedQuestion(category="CS_FUNDAMENTAL", question="DB 트랜잭션 격리수준?"),
-                GeneratedQuestion(category="PROJECT_DEEP_DIVE", question="결제 outbox 어떻게 보장?"),
+                GeneratedQuestion(
+                    category="CS_FUNDAMENTAL", question="DB 트랜잭션 격리수준?"
+                ),
+                GeneratedQuestion(
+                    category="PROJECT_DEEP_DIVE", question="결제 outbox 어떻게 보장?"
+                ),
             ]
         )
     )
@@ -237,25 +241,78 @@ async def test_consumer_injects_initial_rag_chunks_when_available():
                     "summary": "outbox 구현",
                     "techStack": ["Spring"],
                     "markdown": "transactional publisher",
-                }
+                },
+                {
+                    "documentId": 2,
+                    "sourceType": "RESUME",
+                    "summary": "백엔드 3년",
+                    "techStack": ["Java"],
+                    "markdown": "결제 시스템",
+                },
             ],
             "initialQuestionCount": 1,
             "maxQuestions": 5,
         }
     )
     await consumer.handle(_StubMessage(body))
 
+    # 문서 2개 → Retrieve 경로(RAG 검색 발동)
     embedder.embed.assert_awaited_once()
     call = core.search_embeddings.await_args
     assert call.kwargs["query_embedding"] == [0.1, 0.2]
-    assert call.kwargs["document_ids"] == [1]
+    assert call.kwargs["document_ids"] == [1, 2]
     assert call.kwargs["top_k"] == 20  # candidate_k (리랭크 후보 수)
     assert call.kwargs["query_text"]  # 하이브리드 검색: 쿼리 텍스트 동봉
     context = generator.generate.await_args.kwargs["context"]
     assert "Outbox table uses status and retry count" in context
     assert "outbox 구현" in context
 
 
+@pytest.mark.asyncio
+async def test_single_document_skips_rag_plan_mode():
+    # 문서 1개 → PLAN: markdown 이 이미 컨텍스트에 있으니 검색 생략
+    generator = MagicMock()
+    generator.generate = AsyncMock(return_value=GeneratedQuestionPool(questions=[]))
+    publisher = MagicMock()
+    publisher.publish = AsyncMock()
+    core = MagicMock()
+    core.search_embeddings = AsyncMock()
+    embedder = MagicMock()
+    embedder.embed = AsyncMock()
+
+    consumer = QuestionsConsumer(
+        generator=generator,
+        publisher=publisher,
+        idempotency=LruIdempotencyStore(max_size=10),
+        callback_routing_key="callback.questions",
+        core_client=core,
+        embedder=embedder,
+    )
+    body = _envelope(
+        {
+            "sessionId": 1,
+            "mode": "TECHNICAL",
+            "jobCategory": "BACKEND",
+            "documents": [
+                {
+                    "documentId": 1,
+                    "sourceType": "RESUME",
+                    "summary": "백엔드",
+                    "techStack": ["Java"],
+                    "markdown": "결제 시스템 구현",
+                }
+            ],
+            "maxQuestions": 5,
+        }
+    )
+    await consumer.handle(_StubMessage(body))
+
+    core.search_embeddings.assert_not_awaited()
+    embedder.embed.assert_not_awaited()
+    context = generator.generate.await_args.kwargs["context"]
+    assert "결제 시스템 구현" in context  # base context 사용
+
+
 @pytest.mark.asyncio
 async def test_consumer_falls_back_to_document_context_when_rag_fails():
     generator = MagicMock()
@@ -327,3 +384,23 @@ def test_build_context_joins_doc_blocks():
     assert "본문1" in text
     assert "문서 #2 (REPOSITORY)" in text
     assert "readme" in text
+
+
+def test_generated_question_parses_evidence_and_signal():
+    from ai_server.model.messages.questions import GeneratedQuestion
+
+    q = GeneratedQuestion.model_validate(
+        {
+            "category": "PROJECT_DEEP_DIVE",
+            "question": "결제 동시성 어떻게 처리했나요?",
+            "targetEvidence": "결제 시스템에서 분산락 도입",
+            "expectedSignal": "동시성 제어 trade-off 이해",
+        }
+    )
+    assert q.target_evidence == "결제 시스템에서 분산락 도입"
+    assert q.expected_signal == "동시성 제어 trade-off 이해"
+
+    # 하위호환: 신규 필드 없어도 기본값
+    q2 = GeneratedQuestion.model_validate({"category": "BEHAVIORAL", "question": "Q?"})
+    assert q2.target_evidence == ""
+    assert q2.expected_signal == ""

backend/openapi.json

@@ -2517,6 +2517,12 @@
           "createdAt" : {
             "type" : "string",
             "format" : "date-time"
+          },
+          "category" : {
+            "type" : "string"
+          },
+          "targetEvidence" : {
+            "type" : "string"
           }
         }
       },

backend/src/main/java/com/stackup/stackup/session/application/QuestionsCallbackService.java

@@ -86,7 +86,10 @@ private void applyInitialQuestion(InterviewSession session, QuestionsCallbackPay
                 session.getId(), questions.size() - 1);
         }
         InterviewMessage message = messageRepository.save(
-            InterviewMessage.interviewer(session, 1, first.question())
+            InterviewMessage.interviewer(
+                session, 1, first.question(),
+                first.category(), first.targetEvidence(), first.expectedSignal()
+            )
         );
         session.incrementQuestionCount();
         events.publishEvent(new QuestionPersistedEvent(

backend/src/main/java/com/stackup/stackup/session/application/dto/MessageResult.java

@@ -18,7 +18,10 @@ public record MessageResult(
     Double ttsDurationSec,
     Long parentMessageId,
     MessageStatus status,
-    Instant createdAt
+    Instant createdAt,
+    String category,
+    String targetEvidence
+    // expectedSignal 은 의도적으로 제외 — 정답 유출 방지(라이브 비노출).
 ) {
     public static MessageResult of(InterviewMessage m) {
         return new MessageResult(
@@ -33,7 +36,9 @@ public static MessageResult of(InterviewMessage m) {
             m.getTtsDurationSec(),
             m.getParentMessage() == null ? null : m.getParentMessage().getId(),
             m.getStatus(),
-            m.getCreatedAt()
+            m.getCreatedAt(),
+            m.getCategory(),
+            m.getTargetEvidence()
         );
     }
 }

backend/src/main/java/com/stackup/stackup/session/application/dto/QuestionsCallbackPayload.java

@@ -22,7 +22,9 @@ public boolean isFollowup() {
 
     public record GeneratedQuestion(
         String category,
-        String question
+        String question,
+        String targetEvidence,   // 질문 근거(자료 인용). 라이브 노출.
+        String expectedSignal    // 좋은 답이 드러낼 것. 내부 저장만(라이브 비노출).
     ) {
     }
 

backend/src/main/java/com/stackup/stackup/session/domain/InterviewMessage.java

@@ -79,6 +79,16 @@ public class InterviewMessage extends BaseTimeEntity {
     @Column(name = "idempotency_key", length = 64)
     private String idempotencyKey;
 
+    // 질문 메타데이터 (초기 질문에만 채워짐). category/targetEvidence 는 노출, expectedSignal 은 내부용.
+    @Column(length = 30)
+    private String category;
+
+    @Column(name = "target_evidence", columnDefinition = "text")
+    private String targetEvidence;
+
+    @Column(name = "expected_signal", columnDefinition = "text")
+    private String expectedSignal;
+
     private InterviewMessage(InterviewSession session, Integer sequenceNumber, MessageRole role,
                              String content, InterviewMessage parentMessage,
                              MessageStatus initialStatus, String idempotencyKey) {
@@ -101,8 +111,17 @@ private InterviewMessage(InterviewSession session, Integer sequenceNumber, Messa
     }
 
     public static InterviewMessage interviewer(InterviewSession session, int seq, String content) {
+        return interviewer(session, seq, content, null, null, null);
+    }
+
+    public static InterviewMessage interviewer(InterviewSession session, int seq, String content,
+                                               String category, String targetEvidence,
+                                               String expectedSignal) {
         InterviewMessage m = new InterviewMessage(session, seq, MessageRole.INTERVIEWER, content, null,
             MessageStatus.CREATED, null);
+        m.category = category;
+        m.targetEvidence = targetEvidence;
+        m.expectedSignal = expectedSignal;
         m.markTtsPending();
         return m;
     }

backend/src/main/java/com/stackup/stackup/session/presentation/dto/MessageResponse.java

@@ -18,7 +18,10 @@ public record MessageResponse(
     Double ttsDurationSec,
     Long parentMessageId,
     MessageStatus status,
-    Instant createdAt
+    Instant createdAt,
+    String category,
+    String targetEvidence
+    // expectedSignal 은 노출하지 않음 — 자기연습 시 정답 유출 방지.
 ) {
     public static MessageResponse from(MessageResult r) {
         return new MessageResponse(
@@ -33,7 +36,9 @@ public static MessageResponse from(MessageResult r) {
             r.ttsDurationSec(),
             r.parentMessageId(),
             r.status(),
-            r.createdAt()
+            r.createdAt(),
+            r.category(),
+            r.targetEvidence()
         );
     }
 }

backend/src/main/resources/db/migration/V9__add_question_metadata_to_interview_messages.sql

@@ -0,0 +1,12 @@
+-- =============================================================================
+-- 초기 질문 메타데이터: category / target_evidence / expected_signal
+-- =============================================================================
+-- 근거 기반 질문을 위해 AI 가 생성한 메타데이터를 보관한다.
+-- - category: 질문 유형 (CS_FUNDAMENTAL/PROJECT_DEEP_DIVE/TECH_CHOICE/BEHAVIORAL). 느슨 결합(CHECK 없음).
+-- - target_evidence: 질문 근거(자료 인용). 라이브 노출.
+-- - expected_signal: 좋은 답이 드러낼 것. 내부용(라이브 비노출).
+-- 모두 nullable — 기존 행/꼬리질문은 NULL.
+ALTER TABLE interview_messages
+    ADD COLUMN category VARCHAR(30),
+    ADD COLUMN target_evidence TEXT,
+    ADD COLUMN expected_signal TEXT;

backend/src/test/java/com/stackup/stackup/session/application/QuestionsCallbackServiceTest.java

@@ -45,8 +45,9 @@ class QuestionsCallbackServiceTest {
     void apply_poolCallbackStoresOnlyInitialQuestionAndPushesSse() {
         InterviewSession session = sessionFixture(11L, SessionStatus.READY);
         QuestionsCallbackEnvelope env = poolEnvelope(11L, List.of(
-            new GeneratedQuestion("INTRO", "Introduce yourself"),
-            new GeneratedQuestion("TECH", "JPA?")
+            new GeneratedQuestion("PROJECT_DEEP_DIVE", "Introduce yourself",
+                "이력서: 결제 시스템", "협업/문제해결 깊이"),
+            new GeneratedQuestion("TECH", "JPA?", null, null)
         ));
         when(processedMessageRepository.existsById("m-1")).thenReturn(false);
         when(sessionRepository.findById(11L)).thenReturn(Optional.of(session));
@@ -62,6 +63,10 @@ void apply_poolCallbackStoresOnlyInitialQuestionAndPushesSse() {
         verify(messageRepository).save(messageCaptor.capture());
         assertThat(messageCaptor.getValue().getContent()).isEqualTo("Introduce yourself");
         assertThat(messageCaptor.getValue().getSequenceNumber()).isEqualTo(1);
+        // 질문 메타데이터가 첫 질문에서 영속됨
+        assertThat(messageCaptor.getValue().getCategory()).isEqualTo("PROJECT_DEEP_DIVE");
+        assertThat(messageCaptor.getValue().getTargetEvidence()).isEqualTo("이력서: 결제 시스템");
+        assertThat(messageCaptor.getValue().getExpectedSignal()).isEqualTo("협업/문제해결 깊이");
 
         ArgumentCaptor<Object> ev = ArgumentCaptor.forClass(Object.class);
         verify(events, atLeastOnce()).publishEvent(ev.capture());
@@ -108,7 +113,8 @@ void apply_followupAutoEndsSessionAtMaxQuestions() {
 
     @Test
     void apply_skipsDuplicateMessageId() {
-        QuestionsCallbackEnvelope env = poolEnvelope(11L, List.of(new GeneratedQuestion("X", "Q")));
+        QuestionsCallbackEnvelope env =
+            poolEnvelope(11L, List.of(new GeneratedQuestion("X", "Q", null, null)));
         when(processedMessageRepository.existsById("m-1")).thenReturn(true);
 
         service.apply(env);

frontend/src/features/interview/ui/live/QuestionBubble.tsx

@@ -1,10 +1,35 @@
 import type { Message } from '@/domain/session'
+import { StatusBadge } from '@/shared/ui/StatusBadge'
+
+const CATEGORY_LABEL: Record<string, string> = {
+  CS_FUNDAMENTAL: 'CS 기초',
+  PROJECT_DEEP_DIVE: '프로젝트 심화',
+  TECH_CHOICE: '기술 선택',
+  BEHAVIORAL: '인성·행동',
+}
 
 export function QuestionBubble({ message }: { message: Message }) {
+  const categoryLabel = message.category
+    ? (CATEGORY_LABEL[message.category] ?? message.category)
+    : null
+  const hasMeta = Boolean(categoryLabel || message.targetEvidence)
+
   return (
     <div className="flex justify-start">
-      <div className="max-w-[80%] whitespace-pre-wrap rounded-lg rounded-tl-sm bg-surface-raised px-4 py-3 text-body text-fg shadow-sm">
-        {message.content}
+      <div className="flex max-w-[80%] flex-col gap-1.5">
+        {hasMeta && (
+          <div className="flex flex-wrap items-center gap-2">
+            {categoryLabel && <StatusBadge tone="info">{categoryLabel}</StatusBadge>}
+            {message.targetEvidence && (
+              <span className="text-caption text-fg-muted">
+                근거: {message.targetEvidence}
+              </span>
+            )}
+          </div>
+        )}
+        <div className="whitespace-pre-wrap rounded-lg rounded-tl-sm bg-surface-raised px-4 py-3 text-body text-fg shadow-sm">
+          {message.content}
+        </div>
       </div>
     </div>
   )

frontend/src/shared/api/generated.ts

@@ -809,6 +809,8 @@ export interface components {
             status?: "CREATED" | "COMPLETED" | "FAILED";
             /** Format: date-time */
             createdAt?: string;
+            category?: string;
+            targetEvidence?: string;
         };
         VoiceStreamBeginResponse: {
             /** Format: int64 */