feat(feedback): 답변평가 재활용 + 근거 채점 + 피드백 화면 본구현

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

@@ -15,6 +15,11 @@
     "  - improvement_keywords: 다음 면접에서 채울 키워드 5~10개 (짧은 명사구).\n"
     "- 평가 원칙:\n"
     "  - 단일 답변보다 시퀀스 흐름을 우선 고려 (꼬리질문 대응의 깊이가 중요).\n"
+    "  - 전사에 답변별 '답변평가'(specificity/logic/structure/correctness)가 붙어 있으면 "
+    "이를 **근거로 종합**하세요. 무시하고 처음부터 재평가하지 말고 그 점수들을 누적·일관되게 반영합니다.\n"
+    "  - 점수를 매기기 **전에** 강점/약점의 근거를 먼저 정리한 뒤 점수를 산정하세요(즉흥 점수 금지).\n"
+    "  - strengths_summary·weaknesses_summary 에는 **어느 질문/순간 때문인지 구체적으로** 적으세요 "
+    "(예: 'Q3에서 동시성 처리를 RDB 락으로만 설명해 분산 환경 고려가 빠짐').\n"
     "  - 답변이 짧거나 비어 있으면 해당 점수는 낮게 또는 null.\n"
     "  - 컨텍스트 청크(분석 문서 일부) 가 있다면 사실 검증에만 활용 (직접 인용 X).\n"
     "- 응답은 반드시 지정된 JSON 스키마를 따릅니다."

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

@@ -171,10 +171,26 @@ def _build_transcript(messages: list[FeedbackMessageItem]) -> str:
             if m.role == "INTERVIEWER"
             else ("지원자" if m.role == "INTERVIEWEE" else m.role)
         )
-        lines.append(f"[{m.sequence_number}] {speaker}: {m.content}")
+        line = f"[{m.sequence_number}] {speaker}: {m.content}"
+        if m.role == "INTERVIEWEE" and m.evaluation is not None:
+            line += f"\n    └ 답변평가: {_format_evaluation(m.evaluation)}"
+        lines.append(line)
     return "\n".join(lines)
 
 
+def _format_evaluation(e) -> str:
+    parts: list[str] = []
+    if e.specificity is not None:
+        parts.append(f"specificity={e.specificity:g}")
+    if e.logic is not None:
+        parts.append(f"logic={e.logic:g}")
+    if e.structure:
+        parts.append(f"structure={e.structure}")
+    if e.correctness is not None:
+        parts.append(f"correctness={e.correctness:g}")
+    return ", ".join(parts) if parts else "(없음)"
+
+
 def _build_voice_analysis_summary(summary: VoiceAnalysisSummary | None) -> str:
     if summary is None:
         return "No voice analysis summary was provided."

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

@@ -87,6 +87,7 @@ async def handle(self, message: AbstractIncomingMessage) -> None:
                 session_id=req.session_id,
                 kind="FOLLOWUP",
                 parent_message_id=req.parent_message_id,
+                answer_message_id=req.answer_message_id,
                 followup_question=result.followup_question,
                 answer_evaluation=result.answer_evaluation,
             )

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

@@ -7,6 +7,17 @@
 InterviewMode = Literal["PERSONALITY", "TECHNICAL", "INTEGRATED"]
 
 
+class MessageEvaluation(BaseModel):
+    """답변별 평가 (꼬리질문 단계에서 채점·영속된 값). 피드백에서 롤업 재활용."""
+
+    model_config = camel_config()
+
+    specificity: float | None = None
+    logic: float | None = None
+    structure: str | None = None
+    correctness: float | None = None
+
+
 class FeedbackMessageItem(BaseModel):
     """세션 시퀀스 한 줄 (Core 가 통째로 동봉)."""
 
@@ -17,6 +28,8 @@ class FeedbackMessageItem(BaseModel):
     role: Literal["INTERVIEWER", "INTERVIEWEE", "SYSTEM"]
     content: str
     parent_message_id: int | None = None
+    # 답변(INTERVIEWEE) 메시지에만 채워짐. 피드백 종합 채점의 근거.
+    evaluation: MessageEvaluation | None = None
 
 
 class VoiceAnalysisSummary(BaseModel):

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

@@ -51,5 +51,6 @@ class FollowupCallbackPayload(BaseModel):
     session_id: int
     kind: Literal["FOLLOWUP"] = "FOLLOWUP"
     parent_message_id: int
+    answer_message_id: int  # 평가가 달릴 답변 메시지 (Core 가 평가 영속에 사용)
     followup_question: str
     answer_evaluation: AnswerEvaluation | None = None

ai/tests/test_feedback_consumer.py

@@ -234,3 +234,29 @@ async def test_consumer_idempotent_skip():
     await consumer.handle(_StubMessage(_envelope()))
     generator.generate.assert_not_awaited()
     publisher.publish.assert_not_awaited()
+
+
+def test_build_transcript_annotates_interviewee_evaluation():
+    from ai_server.messaging.consumers.feedback_consumer import _build_transcript
+    from ai_server.model.messages.feedback import FeedbackMessageItem, MessageEvaluation
+
+    msgs = [
+        FeedbackMessageItem(
+            id=1, sequence_number=1, role="INTERVIEWER", content="질문?"
+        ),
+        FeedbackMessageItem(
+            id=2,
+            sequence_number=2,
+            role="INTERVIEWEE",
+            content="답변.",
+            evaluation=MessageEvaluation(
+                specificity=2.0, logic=3.0, structure="PARTIAL_STAR", correctness=1.0
+            ),
+        ),
+    ]
+    out = _build_transcript(msgs)
+    assert "답변평가:" in out
+    assert "specificity=2" in out
+    assert "correctness=1" in out
+    # 면접관 줄엔 평가 주석 없음
+    assert out.splitlines()[0].endswith("질문?")

ai/tests/test_followup_consumer.py

@@ -258,3 +258,27 @@ async def test_consumer_passes_parent_category_and_history_to_generator():
     assert kwargs["parent_category"] == "PROJECT_DEEP_DIVE"
     assert "이전 질문" in kwargs["history"]
     assert "면접관:" in kwargs["history"]
+
+
+@pytest.mark.asyncio
+async def test_callback_includes_answer_message_id():
+    generator = MagicMock()
+    generator.generate = AsyncMock(
+        return_value=FollowupResult(
+            followup_question="Q",
+            answer_evaluation=AnswerEvaluation(
+                specificity=2.0, logic=2.0, structure="NONE"
+            ),
+        )
+    )
+    publisher = MagicMock()
+    publisher.publish = AsyncMock()
+    consumer = FollowupConsumer(
+        generator=generator,
+        publisher=publisher,
+        idempotency=LruIdempotencyStore(max_size=10),
+        callback_routing_key="callback.questions",
+    )
+    await consumer.handle(_StubMessage(_envelope()))
+    payload = publisher.publish.await_args.kwargs["payload"]
+    assert payload.answer_message_id == 502  # _envelope 의 answerMessageId

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

@@ -114,6 +114,9 @@ private void applyFollowup(InterviewSession session, QuestionsCallbackPayload pa
             ? null
             : messageRepository.findById(payload.parentMessageId()).orElse(null);
 
+        // 답변 평가를 해당 답변 메시지에 영속 (피드백 롤업 재활용). 평가/대상 없으면 skip.
+        recordAnswerEvaluation(payload);
+
         long currentMsgs = messageRepository.countBySession_Id(session.getId());
         int nextSeq = (int) currentMsgs + 1;
 
@@ -159,6 +162,18 @@ private void applyFollowup(InterviewSession session, QuestionsCallbackPayload pa
     public record SessionStateNotice(Long sessionId, String status, String reason) {
     }
 
+    private void recordAnswerEvaluation(QuestionsCallbackPayload payload) {
+        QuestionsCallbackPayload.AnswerEvaluation eval = payload.answerEvaluation();
+        if (eval == null || payload.answerMessageId() == null) {
+            return;
+        }
+        messageRepository.findById(payload.answerMessageId()).ifPresent(answer -> {
+            answer.recordAnswerEvaluation(
+                eval.specificity(), eval.logic(), eval.structure(), eval.correctness());
+            messageRepository.save(answer);
+        });
+    }
+
     private boolean isProcessed(String messageId) {
         if (messageId == null || messageId.isBlank()) {
             return false;

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

@@ -6,6 +6,7 @@
 import com.stackup.stackup.common.messaging.MessageContext;
 import com.stackup.stackup.common.messaging.RabbitMessagePublisher;
 import com.stackup.stackup.session.application.dto.GenerateFeedbackPayload;
+import com.stackup.stackup.session.application.dto.GenerateFeedbackPayload.MessageEvaluation;
 import com.stackup.stackup.session.application.dto.GenerateFeedbackPayload.MessageItem;
 import com.stackup.stackup.session.application.dto.GenerateFeedbackPayload.VoiceAnalysisSummary;
 import com.stackup.stackup.session.application.event.SessionEndedEvent;
@@ -89,12 +90,24 @@ public void onSessionEnded(SessionEndedEvent event) {
 
     private MessageItem toItem(InterviewMessage m) {
         Long parentId = m.getParentMessage() == null ? null : m.getParentMessage().getId();
+        MessageEvaluation evaluation = null;
+        boolean hasEval = m.getAnswerSpecificity() != null || m.getAnswerLogic() != null
+            || m.getAnswerStructure() != null || m.getAnswerCorrectness() != null;
+        if (m.getRole() == com.stackup.stackup.session.domain.MessageRole.INTERVIEWEE && hasEval) {
+            evaluation = new MessageEvaluation(
+                m.getAnswerSpecificity(),
+                m.getAnswerLogic(),
+                m.getAnswerStructure(),
+                m.getAnswerCorrectness()
+            );
+        }
         return new MessageItem(
             m.getId(),
             m.getSequenceNumber(),
             m.getRole().name(),
             m.getContent(),
-            parentId
+            parentId,
+            evaluation
         );
     }
 

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

@@ -21,7 +21,16 @@ public record MessageItem(
         Integer sequenceNumber,
         String role,
         String content,
-        Long parentMessageId
+        Long parentMessageId,
+        MessageEvaluation evaluation   // INTERVIEWEE 답변에만(없으면 null)
+    ) {
+    }
+
+    public record MessageEvaluation(
+        Double specificity,
+        Double logic,
+        String structure,
+        Double correctness
     ) {
     }
 

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

@@ -8,7 +8,8 @@ public record QuestionsCallbackPayload(
     Long sessionId,
     String kind,                       // POOL | FOLLOWUP
     List<GeneratedQuestion> questions, // POOL legacy kind: initial question result
-    Long parentMessageId,              // FOLLOWUP 시 사용
+    Long parentMessageId,              // FOLLOWUP 시 사용 (직전 질문)
+    Long answerMessageId,              // FOLLOWUP 시 평가 대상 답변 메시지
     String followupQuestion,           // FOLLOWUP 시 사용
     AnswerEvaluation answerEvaluation  // FOLLOWUP 시 옵션
 ) {

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

@@ -89,6 +89,19 @@ public class InterviewMessage extends BaseTimeEntity {
     @Column(name = "expected_signal", columnDefinition = "text")
     private String expectedSignal;
 
+    // 답변 평가 (INTERVIEWEE 메시지에만 채워짐). 꼬리질문 단계 채점값 → 피드백 롤업에 재사용.
+    @Column(name = "answer_specificity")
+    private Double answerSpecificity;
+
+    @Column(name = "answer_logic")
+    private Double answerLogic;
+
+    @Column(name = "answer_structure", length = 20)
+    private String answerStructure;
+
+    @Column(name = "answer_correctness")
+    private Double answerCorrectness;
+
     private InterviewMessage(InterviewSession session, Integer sequenceNumber, MessageRole role,
                              String content, InterviewMessage parentMessage,
                              MessageStatus initialStatus, String idempotencyKey) {
@@ -155,6 +168,15 @@ public void markStatus(MessageStatus newStatus) {
         }
     }
 
+    // 꼬리질문 콜백의 답변 평가를 이 답변 메시지에 기록 (피드백 재활용용).
+    public void recordAnswerEvaluation(Double specificity, Double logic,
+                                       String structure, Double correctness) {
+        this.answerSpecificity = specificity;
+        this.answerLogic = logic;
+        this.answerStructure = structure;
+        this.answerCorrectness = correctness;
+    }
+
     public void attachAudio(String audioFilePath) {
         this.audioFilePath = audioFilePath;
     }

backend/src/main/resources/db/migration/V10__add_answer_evaluation_to_interview_messages.sql

@@ -0,0 +1,10 @@
+-- =============================================================================
+-- 답변 평가 영속 (꼬리질문 단계 채점값) — 피드백에서 롤업 재활용
+-- =============================================================================
+-- 답변(INTERVIEWEE) 메시지에만 채워진다. 모두 nullable.
+-- specificity/logic/correctness: 0~5, structure: STAR enum 문자열.
+ALTER TABLE interview_messages
+    ADD COLUMN answer_specificity DOUBLE PRECISION,
+    ADD COLUMN answer_logic DOUBLE PRECISION,
+    ADD COLUMN answer_structure VARCHAR(20),
+    ADD COLUMN answer_correctness DOUBLE PRECISION;

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

@@ -123,16 +123,46 @@ void apply_skipsDuplicateMessageId() {
         verify(messageRepository, never()).save(any(InterviewMessage.class));
     }
 
+    @Test
+    void apply_followupPersistsAnswerEvaluationOntoAnswerMessage() {
+        InterviewSession session = sessionFixture(11L, SessionStatus.IN_PROGRESS);
+        InterviewMessage answer =
+            InterviewMessage.interviewee(session, 2, "내 답변", null, "idem-1");
+        ReflectionTestUtils.setField(answer, "id", 600L);
+
+        QuestionsCallbackPayload payload = new QuestionsCallbackPayload(
+            11L, "FOLLOWUP", null, 500L, 600L, "다음 질문?",
+            new QuestionsCallbackPayload.AnswerEvaluation(2.0, 3.0, "PARTIAL_STAR", 1.5)
+        );
+        QuestionsCallbackEnvelope env = new QuestionsCallbackEnvelope(
+            "m-eval", "callback.questions", "1", "t", null, "ai", payload, null);
+
+        when(processedMessageRepository.existsById("m-eval")).thenReturn(false);
+        when(sessionRepository.findById(11L)).thenReturn(Optional.of(session));
+        when(messageRepository.findById(500L)).thenReturn(Optional.empty());
+        when(messageRepository.findById(600L)).thenReturn(Optional.of(answer));
+        when(messageRepository.countBySession_Id(11L)).thenReturn(2L);
+        when(messageRepository.save(any(InterviewMessage.class)))
+            .thenAnswer(inv -> inv.getArgument(0));
+
+        service.apply(env);
+
+        assertThat(answer.getAnswerSpecificity()).isEqualTo(2.0);
+        assertThat(answer.getAnswerLogic()).isEqualTo(3.0);
+        assertThat(answer.getAnswerStructure()).isEqualTo("PARTIAL_STAR");
+        assertThat(answer.getAnswerCorrectness()).isEqualTo(1.5);
+    }
+
     private QuestionsCallbackEnvelope poolEnvelope(Long sessionId, List<GeneratedQuestion> questions) {
         QuestionsCallbackPayload payload = new QuestionsCallbackPayload(
-            sessionId, "POOL", questions, null, null, null
+            sessionId, "POOL", questions, null, null, null, null
         );
         return new QuestionsCallbackEnvelope("m-1", "callback.questions", "1", "t", null, "ai", payload, null);
     }
 
     private QuestionsCallbackEnvelope followupEnvelope(Long sessionId, Long parentId, String followup) {
         QuestionsCallbackPayload payload = new QuestionsCallbackPayload(
-            sessionId, "FOLLOWUP", null, parentId, followup, null
+            sessionId, "FOLLOWUP", null, parentId, null, followup, null
         );
         return new QuestionsCallbackEnvelope("m-2", "callback.questions", "1", "t", null, "ai", payload, null);
     }

frontend/src/features/feedback/api/feedbackApi.ts

@@ -0,0 +1,9 @@
+import { apiClient } from '@/shared/api'
+import type { components } from '@/shared/api/generated'
+
+type S = components['schemas']
+export type Feedback = S['FeedbackResponse']
+
+export async function getFeedback(sessionId: number): Promise<Feedback> {
+  return (await apiClient.get<Feedback>(`/api/sessions/${sessionId}/feedback`)).data
+}

frontend/src/features/feedback/index.ts

@@ -0,0 +1,3 @@
+export { FeedbackReport } from './ui/FeedbackReport'
+export { useFeedback, isFeedbackPending, feedbackKeys } from './model/useFeedback'
+export type { Feedback } from './api/feedbackApi'

frontend/src/features/feedback/model/useFeedback.ts

@@ -0,0 +1,23 @@
+import { useQuery } from '@tanstack/react-query'
+import { isApiError } from '@/shared/api'
+import { getFeedback } from '../api/feedbackApi'
+
+export const feedbackKeys = {
+  all: ['feedback'] as const,
+  detail: (sessionId: number) => [...feedbackKeys.all, sessionId] as const,
+}
+
+// 세션 종료 직후엔 피드백이 비동기 생성 중이라 아직 없음(FEEDBACK_NOT_READY/404).
+// 이 경우는 에러가 아니라 "생성 중"이므로 일정 횟수까지 polling 한다.
+export function isFeedbackPending(err: unknown): boolean {
+  return isApiError(err) && (err.code === 'FEEDBACK_NOT_READY' || err.status === 404)
+}
+
+export function useFeedback(sessionId: number) {
+  return useQuery({
+    queryKey: feedbackKeys.detail(sessionId),
+    queryFn: () => getFeedback(sessionId),
+    retry: (count, err) => isFeedbackPending(err) && count < 40,
+    retryDelay: 3000,
+  })
+}