happybell80
0a967ce7c1
- 검색 경로에 벡터+키워드+그래프 RRF 합산 명시 - Storage에 AGE 추가 - Question Types 4종 섹션 추가 - Response Shape를 Pydantic 모델 기준으로 갱신 Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
61 lines
2.8 KiB
Markdown
61 lines
2.8 KiB
Markdown
---
|
|
name: companyx-rag
|
|
description: Use this skill when a Company X user asks for answers grounded in Company X internal documents, especially for evidence checks, internal-material summaries, program explanations, or document-based fact verification.
|
|
---
|
|
|
|
# Company X RAG
|
|
|
|
Use this skill only for Company X users and only when the answer should be grounded in Company X internal documents.
|
|
|
|
## Trigger
|
|
|
|
- The user is authenticated as a Company X member.
|
|
- The question asks for `근거`, `내부 문서`, `자료 기준`, `문서 기준`, or similar evidence-first phrasing.
|
|
- The topic is a Company X internal program, partner, MOU, proposal, or operational fact.
|
|
|
|
## Do
|
|
|
|
- Search Company X internal documents before using general web knowledge.
|
|
- Answer in this order:
|
|
1. direct answer
|
|
2. evidence documents
|
|
3. short evidence summary
|
|
- If documents conflict, say they conflict.
|
|
- If documents are missing, say the evidence is insufficient.
|
|
|
|
## Do Not
|
|
|
|
- Do not guess missing internal facts.
|
|
- Do not answer from general web knowledge first when the user expects internal grounding.
|
|
- Do not dump long raw excerpts without a short explanation.
|
|
- Do not expose Company X internal grounding to non-Company X users.
|
|
|
|
## Current Operating Path
|
|
|
|
- Team boundary: `79441171-3951-4870-beb8-916d07fe8be5`
|
|
- Retrieval service: `skill-rag-file` (`/api/search`, `search_mode=hybrid`)
|
|
- 검색 경로: 벡터(PGVector cosine) + 키워드(TSVECTOR + GIN) + 그래프(Apache AGE) → RRF 합산
|
|
- Embedding: Gemini Embedding 2 (`gemini-embedding-2-preview`), `768d`, skill-embedding 게이트웨이 경유
|
|
- Storage: PostgreSQL (pgvector + AGE) 단일 운영. ChromaDB는 레거시(운영 비사용).
|
|
- 청킹: 텍스트 추출 후 문자 단위 분할 (chunk_size=1000, overlap=200)
|
|
- 인덱싱 대상: 200개 파일 (`latest_200_companyx.txt`), DB 기준 `team_document` 1,172건 / `team_document_chunk` 3,095건
|
|
- NAS 원본: `/mnt/nas/workspace/6.Company X` (53,249 파일)
|
|
|
|
## Question Types
|
|
|
|
- `fact_check`: 사실 확인 (기본값) — "근거 있어?", "협력 관계야?"
|
|
- `explanatory`: 설명 요청 — "뭐야?", "설명해줘"
|
|
- `quantitative`: 수치 질문 — "몇 개야?", "얼마나?"
|
|
- `recap`: 재정리 — "다시 정리해줘", "문서명만"
|
|
|
|
## Response Shape
|
|
|
|
- Pydantic 모델: `CompanyXRAGOutput`
|
|
- `direct_answer`: 1~3문장 답변 (근거 부족 시 빈 문자열)
|
|
- `evidence_docs`: 근거 문서 파일명 목록
|
|
- `failure_reason`: 답변 불가 사유 (`null` 또는 "문서 없음", "단정 불가" 등)
|
|
- 사용자 응답 구조:
|
|
1. Direct answer: `네/아니요/현재는 단정 불가`
|
|
2. Evidence section: 문서명, snippet 요약, chunk/page, 점수(score, vec, kw)
|
|
3. Limitation section (필요 시): `문서 없음`, `문서 간 불일치`, `최신 집계 미확인`
|