효성 RAGAS 구현 구조 한눈에 보기

flowchart LR
  %% Hyosung RAGAS Assessment architecture

  USER["사용자"] --> CHATAPI["FastAPI Chat API
ai_chat_json_router"]
  CHATAPI --> SEARCH["Internal Search Service
hybrid retrieval / RRF"]
  SEARCH --> OPENSEARCH["OpenSearch
BM25 + Vector KNN"]
  OPENSEARCH --> CONTEXTS["Top-K Contexts
source metadata"]
  CONTEXTS --> LLM["LiteLLM / LLM
응답 생성"]
  LLM --> RESPONSE["응답 반환
사용자 지연 없음"]

  RESPONSE --> SAMPLE{"샘플링 정책
0 / 10 / 50 / 100%"}
  SAMPLE -->|제외| SKIP["평가 미수행
메타 로그만 유지"]
  SAMPLE -->|운영 샘플| OPJOB["Operational Eval Job
reference 없음"]
  GOLDEN["Golden Set
reference / reference_contexts"] --> BATCH["Batch Eval Job
4대 지표 전체"]

  subgraph NEW["신규 RAGAS 평가 계층"]
    OPJOB --> QUEUE["Redis / PostgreSQL Queue
status / retry / idempotency"]
    BATCH --> QUEUE
    QUEUE --> WORKER["RAGAS Worker
timeout / retry / cancel"]
    WORKER --> ADAPTER["RagasEvaluator Adapter
RAGAS API 변화 격리"]
    ADAPTER --> JUDGE["Judge LLM
LiteLLM 경유"]
    ADAPTER --> RAGAS["vibrantlabsai/ragas
0.4.x"]
  end

  subgraph METRICS["메트릭 적용 범위"]
    RAGAS --> OP_METRICS["운영 트래픽
faithfulness
answer relevancy"]
    RAGAS --> GOLD_METRICS["Golden Set
faithfulness
answer relevancy
context precision
context recall"]
  end

  subgraph STORAGE["PostgreSQL 저장/집계"]
    OP_METRICS --> EVAL["ragas_evaluation
상태 / 점수 / 메타"]
    GOLD_METRICS --> EVAL
    EVAL --> SCORE["ragas_metric_score
지표별 점수"]
    EVAL --> PAYLOAD["ragas_payload
질의 / 응답 / 컨텍스트"]
    SCORE --> SUMMARY["summary table
materialized view
100만 건 통계 대응"]
  end

  subgraph ADMIN["Admin RAGAS Assessment"]
    SUMMARY --> DASH["실시간 현황
큐 적체 / 실패 / 평균"]
    SUMMARY --> TREND["기간 통계
p5 / p95 / 분포"]
    EVAL --> DRILL["케이스 드릴다운
Judge reasoning"]
    PAYLOAD --> DRILL
    GOLDEN --> GOLDADMIN["Golden Set 관리
import / review / batch"]
    SUMMARY --> ALERT["임계값 알림
cooldown / retry"]
    EVAL --> EXPORT["CSV / JSON Export"]
  end

  classDef existing fill:#e8f2ff,stroke:#3b73b9,color:#202938,stroke-width:1.5px;
  classDef proposed fill:#fff4d8,stroke:#b98500,color:#202938,stroke-width:2px;
  classDef storage fill:#eaf7ea,stroke:#3d8b3d,color:#202938,stroke-width:1.7px;
  classDef admin fill:#f3edff,stroke:#7660c7,color:#202938,stroke-width:1.7px;
  classDef risk fill:#ffe8ed,stroke:#df4f70,color:#202938,stroke-width:1.8px;

  class USER,CHATAPI,SEARCH,OPENSEARCH,CONTEXTS,LLM,RESPONSE existing;
  class SAMPLE,OPJOB,GOLDEN,BATCH,QUEUE,WORKER,ADAPTER,JUDGE,RAGAS,OP_METRICS,GOLD_METRICS proposed;
  class EVAL,SCORE,PAYLOAD,SUMMARY storage;
  class DASH,TREND,DRILL,GOLDADMIN,ALERT,EXPORT admin;
  class SKIP risk;