From 72bb66b8ebe275f1fb0cceb8b3645f65856a15f3 Mon Sep 17 00:00:00 2001
From: happybell80 <happybell80@gmail.com>
Date: Tue, 14 Oct 2025 15:56:28 +0900
Subject: [PATCH] =?UTF-8?q?docs:=20coldmail=20=ED=95=98=EC=9D=B4=EB=B8=8C?=
 =?UTF-8?q?=EB=A6=AC=EB=93=9C=20=ED=95=84=ED=84=B0=20=EA=B5=AC=ED=98=84=20?=
 =?UTF-8?q?=EA=B3=84=ED=9A=8D=20=EC=A0=95=EA=B5=90=ED=99=94?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

- Phase 1: skill-embedding(8515) /embed 엔드포인트, 384차원 명시
- Phase 2: Gemini gemini-2.5-flash-lite 모델명 추가
- Phase 3: coldmail_filter.py:44-52 정규식 교체 위치 명시
- Phase 4: scheduler→services 리팩토링 상세화 (76-263줄, 8단계 워크플로우)
- DB: pgvector extension, coldmail_embedding_clusters 스키마 추가
- 환경변수: SKILL_EMBEDDING_URL 참조 파일 명시

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <noreply@anthropic.com>
---
 ...aude_coldmail_filter_tokenization_issue.md | 30 +++++++++++++------
 1 file changed, 21 insertions(+), 9 deletions(-)

diff --git a/troubleshooting/251014_claude_coldmail_filter_tokenization_issue.md b/troubleshooting/251014_claude_coldmail_filter_tokenization_issue.md
index 684deb3..9ca5be7 100644
--- a/troubleshooting/251014_claude_coldmail_filter_tokenization_issue.md
+++ b/troubleshooting/251014_claude_coldmail_filter_tokenization_issue.md
@@ -41,27 +41,39 @@
 ## 구현 계획 (12시간)
 
 ### Phase 1: KoBERT 임베딩 필터 (4h)
-**신규**: coldmail_embedding_filter.py
-- create_embedding(): skill-embedding (8515)로 임베딩 생성
+**신규**: app/services/coldmail_embedding_filter.py
+- create_embedding(): skill-embedding (8515) /embed 호출, 384차원
 - calculate_similarity(): cosine similarity
 - is_coldmail_by_embedding(): threshold 0.6
+- **환경변수**: SKILL_EMBEDDING_URL (emotion_classifier.py:9, embedding_client.py:14 참고)
 
-**DB**: coldmail_embedding_clusters 테이블 (pgvector extension, VECTOR(384) 필요)
+**DB**: coldmail_embedding_clusters 테이블
+- pgvector extension 설치 필요: CREATE EXTENSION IF NOT EXISTS vector;
+- 컬럼: id, embedding VECTOR(384), label TEXT, created_at TIMESTAMP
 
 ### Phase 2: Gemini LLM 분류기 (2h)
-**신규**: coldmail_llm_classifier.py
+**신규**: app/services/coldmail_llm_classifier.py
 - classify_by_llm(): Gemini zero-shot + 이유 설명
 - JSON 마크다운 블록 전처리
+- **모델**: gemini-2.5-flash-lite (emotion_classifier.py 참고)
 
 ### Phase 3: 하이브리드 통합 (3h)
-**신규**: coldmail_hybrid_filter.py
+**신규**: app/services/coldmail_hybrid_filter.py
 - hybrid_coldmail_filter(): 3단계 순차 실행
 - update_from_feedback(): Slack 버튼 → 3모델 업데이트
+- **기존**: app/services/coldmail_filter.py:44-52 정규식 토큰화 교체
 
-### Phase 4: coldmail_briefing 통합 (1h)
-**수정**: coldmail_briefing.py:121-136
-- is_coldmail() → hybrid_coldmail_filter() 교체
-- 분류 상세 정보 로그 추가
+### Phase 4: 아키텍처 리팩토링 (3h)
+**신규**: app/services/coldmail_briefing_service.py
+- _run_coldmail_briefing() 이동 (scheduler/jobs/coldmail_briefing.py:76-263)
+- 8단계 워크플로우: 시간 계산 → 메일 조회 → 콜드메일 필터 → 첨부파일 처리 → IR 분석 → 가치평가 → Slack Lists → 피드백 버튼
+
+**수정**: app/scheduler/jobs/coldmail_briefing.py
+- register(), _parse_cron(), _run_coldmail_briefing_with_logging() 유지 (1-75줄)
+- is_coldmail() → hybrid_coldmail_filter() 교체 (133줄)
+
+**LangGraph 검토**: 250925_langgraph_vs_n8n_comparison.md
+- 체크포인트 기반 실패 복구 (IR 분석 실패 시 재시도)
 
 ### Phase 5: 테스트 및 검증 (2h)
 **테스트 케이스** (6개):