feat: 피드백 상세화 — 면접관별 상세평가·점수근거 + 종합 서술형 + 학습 방향

ai/src/ai_server/chain/feedback_generation_chain.py

@@ -11,7 +11,7 @@
 from pydantic import BaseModel, Field
 
 from ai_server.chain.prompts.feedback_generation import HUMAN_PROMPT, SYSTEM_PROMPT
-from ai_server.chain.prompts import feedback_panel
+from ai_server.chain.prompts import feedback_panel, feedback_synthesis
 from ai_server.config.settings import Settings
 from ai_server.core.client import CoreClient
 from ai_server.model.messages.feedback import PanelBreakdownItem
@@ -28,6 +28,7 @@ class FeedbackResult(BaseModel):
     strengths_summary: str | None = Field(None)
     weaknesses_summary: str | None = Field(None)
     improvement_keywords: list[str] = Field(default_factory=list)
+    study_plan: list[str] = Field(default_factory=list)
     panel_breakdown: list[PanelBreakdownItem] = Field(default_factory=list)
 
 
@@ -126,9 +127,18 @@ class EvaluatorResult(BaseModel):
     score: float | None = Field(None, description="0~100, 산정 불가 시 null")
     strength: str | None = None
     weakness: str | None = None
+    detail: str | None = Field(None, description="근거·예시 포함 2~4문장 상세 평가")
+    score_rationale: str | None = Field(None, description="점수 근거 한두 문장")
     keywords: list[str] = Field(default_factory=list)
 
 
+class SynthesisResult(BaseModel):
+    strengths_summary: str | None = None
+    weaknesses_summary: str | None = None
+    improvement_keywords: list[str] = Field(default_factory=list)
+    study_plan: list[str] = Field(default_factory=list)
+
+
 @dataclass(frozen=True)
 class _EvaluatorSpec:
     key: str  # 'technical' | 'logic' | 'communication'
@@ -253,6 +263,40 @@ def build_panel_evaluator_chain(
     return prompt | llm | parser
 
 
+def build_feedback_synthesis_chain(
+    settings: Settings, core_client: CoreClient | None = None
+) -> Runnable:
+    """패널 결과를 통합한 종합 서술형 평 + 학습 방향 생성 체인."""
+    from langchain_openai import ChatOpenAI
+
+    parser = PydanticOutputParser(pydantic_object=SynthesisResult)
+    prompt = ChatPromptTemplate.from_messages(
+        [
+            ("system", feedback_synthesis.SYSTEM_PROMPT),
+            ("human", feedback_synthesis.HUMAN_PROMPT),
+        ]
+    ).partial(format_instructions=parser.get_format_instructions())
+
+    callbacks = []
+    if core_client is not None:
+        callbacks.append(
+            CoreAiLogCallback(
+                core_client=core_client,
+                request_type="generate.feedback.synthesis",
+                default_model=settings.llm_pro_model,
+            )
+        )
+
+    llm = ChatOpenAI(
+        model=settings.llm_pro_model,
+        temperature=settings.llm_pro_temperature,
+        api_key=settings.llm_api_key or None,
+        base_url=settings.llm_base_url,
+        callbacks=callbacks,
+    )
+    return prompt | llm | parser
+
+
 def _weighted_overall(pairs: list[tuple[float | None, float]]) -> float | None:
     """(score, weight) 중 score 가 있는 것만 가중평균. 전부 None 이면 None."""
     present = [(s, w) for s, w in pairs if s is not None and w > 0]
@@ -283,8 +327,15 @@ def _dedup_keywords(keywords: list[str], cap: int = 8) -> list[str]:
 class PanelFeedbackGenerator:
     """직군·논리·커뮤니케이션 평가위원을 병렬 호출 → 가중평균 종합. FeedbackGenerator 호환."""
 
-    def __init__(self, chain: Runnable, *, weights: tuple[float, float, float] = (0.5, 0.25, 0.25)) -> None:
+    def __init__(
+        self,
+        chain: Runnable,
+        *,
+        synthesis_chain: Runnable | None = None,
+        weights: tuple[float, float, float] = (0.5, 0.25, 0.25),
+    ) -> None:
         self._chain = chain
+        self._synthesis = synthesis_chain
         self._w_tech, self._w_logic, self._w_comm = weights
 
     async def generate(
@@ -353,18 +404,18 @@ async def generate(
             ]
         )
 
-        note_items = [(spec.label, r.strength) for spec, _, r in domain_results]
-        note_items += [("논리", logic.strength), ("전달", comm.strength)]
-        strengths = _merge_notes(note_items)
-        weak_items = [(spec.label, r.weakness) for spec, _, r in domain_results]
-        weak_items += [("논리", logic.weakness), ("전달", comm.weakness)]
-        weaknesses = _merge_notes(weak_items)
-
-        all_keywords: list[str] = []
-        for _, _, r in domain_results:
-            all_keywords += r.keywords
-        all_keywords += logic.keywords + comm.keywords
-        keywords = _dedup_keywords(all_keywords)
+        # 기계적 병합(synthesis 미설정/실패 시 폴백용).
+        fb_strength_items = [(spec.label, r.strength) for spec, _, r in domain_results]
+        fb_strength_items += [("논리", logic.strength), ("전달", comm.strength)]
+        fb_strengths = _merge_notes(fb_strength_items)
+        fb_weak_items = [(spec.label, r.weakness) for spec, _, r in domain_results]
+        fb_weak_items += [("논리", logic.weakness), ("전달", comm.weakness)]
+        fb_weaknesses = _merge_notes(fb_weak_items)
+        fb_keywords = _dedup_keywords(
+            [k for _, _, r in domain_results for k in r.keywords]
+            + logic.keywords
+            + comm.keywords
+        )
 
         breakdown = [
             PanelBreakdownItem(
@@ -373,6 +424,8 @@ async def generate(
                 score=r.score,
                 strength=r.strength,
                 weakness=r.weakness,
+                detail=r.detail,
+                score_rationale=r.score_rationale,
             )
             for spec, _, r in domain_results
         ]
@@ -384,9 +437,37 @@ async def generate(
                     score=r.score,
                     strength=r.strength,
                     weakness=r.weakness,
+                    detail=r.detail,
+                    score_rationale=r.score_rationale,
                 )
             )
 
+        # 종합 서술형 + 학습 방향(synthesis). 미설정/실패 시 기계적 병합으로 폴백.
+        strengths, weaknesses, keywords, study_plan = fb_strengths, fb_weaknesses, fb_keywords, []
+        if self._synthesis is not None:
+            panel_summary = "\n".join(
+                f"- {b.evaluator}({b.dimension}) "
+                f"{('%g점' % b.score) if b.score is not None else '점수없음'}: "
+                f"{(b.detail or '').strip() or (b.strength or '')}"
+                for b in breakdown
+            )
+            try:
+                syn = await self._synthesis.ainvoke(
+                    {
+                        "job_category": job_category,
+                        "mode": mode,
+                        "panel_summary": panel_summary,
+                        "transcript": transcript,
+                    }
+                )
+                if isinstance(syn, SynthesisResult):
+                    strengths = syn.strengths_summary or fb_strengths
+                    weaknesses = syn.weaknesses_summary or fb_weaknesses
+                    keywords = _dedup_keywords(syn.improvement_keywords) or fb_keywords
+                    study_plan = syn.study_plan or []
+            except Exception as exc:  # noqa: BLE001
+                log.warning("feedback.synthesis.failed", error=str(exc))
+
         return FeedbackResult(
             overall_score=overall,
             technical_accuracy=technical_accuracy,
@@ -395,5 +476,6 @@ async def generate(
             strengths_summary=strengths,
             weaknesses_summary=weaknesses,
             improvement_keywords=keywords,
+            study_plan=study_plan,
             panel_breakdown=breakdown,
         )

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

@@ -14,6 +14,9 @@
     "- 점수를 매기기 전에 강점/약점 근거를 먼저 정리한 뒤 산정한다(즉흥 점수 금지).\n"
     "- strength/weakness 는 각각 한 줄(한국어, 구체적으로). keywords 는 이 축에서 보완할 "
     "개선 키워드 0~4개(짧은 명사구).\n"
+    "- detail: 이 축에 대한 2~4문장 상세 평가. **답변의 구체적 부분을 인용/지목**하고 무엇이 "
+    "왜 좋았는지/아쉬웠는지 근거를 들어 서술한다(추상적 총평 금지).\n"
+    "- score_rationale: 그 점수를 준 핵심 근거(가점/감점 요인)를 한두 문장으로.\n"
     "- 응답은 반드시 지정된 JSON 스키마를 따른다."
 )
 

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

@@ -0,0 +1,23 @@
+# 멀티 면접관 패널 종합 — 평가위원들의 결과를 통합해 서술형 종합평 + 학습 방향 생성.
+
+SYSTEM_PROMPT = (
+    "당신은 면접 평가 패널을 총괄하는 수석 면접관입니다. 여러 평가위원(직군 기술·논리·"
+    "전달 등)의 평가를 받아 지원자에게 줄 **통합 종합 피드백**을 작성하세요.\n"
+    "- 평가위원들의 관점을 단순 나열하지 말고 **하나의 일관된 서술**로 엮습니다.\n"
+    "- strengths_summary: 지원자의 강점을 3~5문장 한국어 문단으로. 어느 영역에서 어떻게 "
+    "드러났는지 구체적으로.\n"
+    "- weaknesses_summary: 보완점을 3~5문장 한국어 문단으로. 무엇이 왜 부족했고 어떤 방향으로 "
+    "개선해야 하는지.\n"
+    "- improvement_keywords: 다음 면접에서 보완할 키워드 5~10개(짧은 명사구).\n"
+    "- study_plan: 구체적 학습 방향/다음 단계 액션 아이템 3~6개. 각 항목은 '무엇을 어떻게'가 "
+    "드러나는 실행 문장(예: 'Redis 분산 락의 SETNX·TTL 옵션을 직접 구현해보며 원자성 보장 원리 정리').\n"
+    "- 평가위원 점수·근거에 모순되지 않게, 과장 없이 사실에 기반해 작성합니다.\n"
+    "- 응답은 반드시 지정된 JSON 스키마를 따릅니다."
+)
+
+HUMAN_PROMPT = (
+    "직군: {job_category} / 면접 모드: {mode}\n\n"
+    "=== 평가위원별 평가 결과 ===\n{panel_summary}\n\n"
+    "=== 면접 전사(참고) ===\n{transcript}\n\n"
+    "{format_instructions}"
+)

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

@@ -109,6 +109,7 @@ async def handle(self, message: AbstractIncomingMessage) -> None:
                 strengths_summary=result.strengths_summary,
                 weaknesses_summary=result.weaknesses_summary,
                 improvement_keywords=result.improvement_keywords,
+                study_plan=result.study_plan,
                 panel_breakdown=result.panel_breakdown,
                 report_s3_key=None,
             )

ai/src/ai_server/messaging/runner.py

@@ -21,6 +21,7 @@
 )
 from ai_server.chain.feedback_generation_chain import (
     PanelFeedbackGenerator,
+    build_feedback_synthesis_chain,
     build_panel_evaluator_chain,
 )
 from ai_server.chain.question_generation_chain import (
@@ -211,8 +212,10 @@ def __init__(self, settings: Settings) -> None:
         )
 
         # 종합 피드백 생성 (US-24) — 멀티 면접관 패널(직군·논리·커뮤 평가위원 병렬 → 가중평균)
+        # + 종합 서술형 평·학습 방향 synthesis.
         feedback_generator = PanelFeedbackGenerator(
-            build_panel_evaluator_chain(settings, core_client=core_client)
+            build_panel_evaluator_chain(settings, core_client=core_client),
+            synthesis_chain=build_feedback_synthesis_chain(settings, core_client=core_client),
         )
         self._feedback_consumer = FeedbackConsumer(
             generator=feedback_generator,

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

@@ -74,6 +74,10 @@ class PanelBreakdownItem(BaseModel):
     score: float | None = None
     strength: str | None = None
     weakness: str | None = None
+    # 구체적 근거·예시(답변 인용)를 담은 2~4문장 상세 평가.
+    detail: str | None = None
+    # 이 점수를 준 근거(감점/가점 요인) 간단 설명.
+    score_rationale: str | None = None
 
 
 class FeedbackCallbackPayload(BaseModel):
@@ -89,6 +93,8 @@ class FeedbackCallbackPayload(BaseModel):
     strengths_summary: str | None = None
     weaknesses_summary: str | None = None
     improvement_keywords: list[str] = Field(default_factory=list)
+    # 패널을 통합한 학습 방향/다음 단계 액션 아이템.
+    study_plan: list[str] = Field(default_factory=list)
     # 평가위원별 분해(패널). 비어 있으면 단일/레거시 경로.
     panel_breakdown: list[PanelBreakdownItem] = Field(default_factory=list)
     report_s3_key: str | None = None

ai/tests/test_feedback_panel.py

@@ -3,8 +3,18 @@
 from ai_server.chain.feedback_generation_chain import (
     EvaluatorResult,
     PanelFeedbackGenerator,
+    SynthesisResult,
 )
 
+
+class _FakeSynthesis:
+    def __init__(self, result: SynthesisResult):
+        self._result = result
+
+    async def ainvoke(self, _v):
+        return self._result
+
+
 # 평가축(dimension_name) 으로 라우팅하는 가짜 체인.
 TECH = "기술 정확도·깊이"
 PERSONALITY = "인성·협업 역량"
@@ -123,6 +133,44 @@ async def test_multi_domain_weighted_by_question_counts():
     assert r.overall_score == 62
 
 
+@pytest.mark.asyncio
+async def test_synthesis_narrative_study_plan_and_breakdown_detail():
+    syn = SynthesisResult(
+        strengths_summary="통합 강점 서술",
+        weaknesses_summary="통합 약점 서술",
+        improvement_keywords=["동시성"],
+        study_plan=["Redis 분산 락 SETNX/TTL 직접 구현"],
+    )
+    gen = PanelFeedbackGenerator(
+        _FakeChain({
+            TECH: EvaluatorResult(score=80, strength="s", detail="상세 평가", score_rationale="근거"),
+            LOGIC: EvaluatorResult(score=60),
+            COMM: EvaluatorResult(score=50),
+        }),
+        synthesis_chain=_FakeSynthesis(syn),
+    )
+    r = await _run_gen(gen)
+    # 종합 서술형(synthesis 결과로 대체)
+    assert r.strengths_summary == "통합 강점 서술"
+    assert r.weaknesses_summary == "통합 약점 서술"
+    assert r.improvement_keywords == ["동시성"]
+    assert r.study_plan == ["Redis 분산 락 SETNX/TTL 직접 구현"]
+    # 평가위원 분해에 detail/score_rationale 포함
+    assert r.panel_breakdown[0].detail == "상세 평가"
+    assert r.panel_breakdown[0].score_rationale == "근거"
+
+
+async def _run_gen(gen):
+    return await gen.generate(
+        job_category="BACKEND",
+        mode="TECHNICAL",
+        total_question_count=3,
+        end_reason="x",
+        transcript="t",
+        rag_context="(none)",
+    )
+
+
 @pytest.mark.asyncio
 async def test_keyword_dedup():
     r = await _run(

backend/openapi.json

@@ -3348,6 +3348,13 @@
               "$ref" : "#/components/schemas/PanelBreakdownItem"
             }
           },
+          "studyPlan" : {
+            "type" : "array",
+            "description" : "Study plan / next-step action items synthesized from the panel.",
+            "items" : {
+              "type" : "string"
+            }
+          },
           "reportFilePath" : {
             "type" : "string",
             "description" : "Stored report path when AI generates a detailed learning guide/report."
@@ -3376,6 +3383,12 @@
           },
           "weakness" : {
             "type" : "string"
+          },
+          "detail" : {
+            "type" : "string"
+          },
+          "scoreRationale" : {
+            "type" : "string"
           }
         }
       },

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

@@ -76,6 +76,7 @@ public void apply(FeedbackCallbackEnvelope envelope) {
             payload.weaknessesSummary(),
             keywordsToJson(payload.improvementKeywords()),
             breakdownToJson(payload.panelBreakdown()),
+            keywordsToJson(payload.studyPlan()),
             payload.reportS3Key()
         );
         try {

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

@@ -14,6 +14,7 @@ public record FeedbackCallbackPayload(
     String strengthsSummary,
     String weaknessesSummary,
     List<String> improvementKeywords,
+    List<String> studyPlan,
     List<PanelBreakdownItem> panelBreakdown,
     String reportS3Key
 ) {

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

@@ -19,6 +19,7 @@ public record FeedbackResult(
     String weaknessesSummary,
     List<String> improvementKeywords,
     List<PanelBreakdownItem> panelBreakdown,
+    List<String> studyPlan,
     String reportFilePath,
     Instant createdAt
 ) {
@@ -37,6 +38,7 @@ public static FeedbackResult of(SessionFeedback f) {
             f.getWeaknessesSummary(),
             parseKeywords(f.getImprovementKeywords()),
             parseBreakdown(f.getPanelBreakdown()),
+            parseKeywords(f.getStudyPlan()),
             f.getReportFilePath(),
             f.getCreatedAt()
         );

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

@@ -9,6 +9,8 @@ public record PanelBreakdownItem(
     String dimension,
     Double score,
     String strength,
-    String weakness
+    String weakness,
+    String detail,          // 근거·예시 포함 상세 평가
+    String scoreRationale   // 점수 근거
 ) {
 }