f9af8dd355
Phase 1.2-B 평가셋 결과 recall 0.788 → 0.750 회귀. 원인: trigram default threshold 0.3이 multi-token 쿼리에서 너무 엄격. 예: '이란 미국 전쟁 글로벌 반응' 같은 5단어 한국어 뉴스 쿼리는 title/ai_summary trigram 매칭이 거의 안 됨. 해결: search_text 시작 시 set_limit(0.15) 호출. - trigram 매칭 더 관대 (recall ↑) - precision은 ORDER BY similarity 가중 합산이 보정 - p95 latency 169ms 여유 충분 (목표 500ms)
152 lines
6.5 KiB
Python
152 lines
6.5 KiB
Python
"""검색 후보 수집 서비스 (Phase 1.2).
|
|
|
|
text(documents FTS + trigram) + vector(documents.embedding → chunks) 후보를
|
|
SearchResult 리스트로 반환.
|
|
|
|
Phase 1.1a: search.py의 _search_text/_search_vector를 이전 (ILIKE 그대로).
|
|
Phase 1.2-B: ILIKE → trigram `%` + `similarity()`. ILIKE 풀 스캔 제거.
|
|
Phase 1.2-B 이후: vector retrieval을 document_chunks 테이블 기반으로 전환.
|
|
"""
|
|
|
|
from __future__ import annotations
|
|
|
|
from typing import TYPE_CHECKING
|
|
|
|
from sqlalchemy import text
|
|
from sqlalchemy.ext.asyncio import AsyncSession
|
|
|
|
from ai.client import AIClient
|
|
|
|
if TYPE_CHECKING:
|
|
from api.search import SearchResult
|
|
|
|
|
|
async def search_text(
|
|
session: AsyncSession, query: str, limit: int
|
|
) -> list["SearchResult"]:
|
|
"""FTS + trigram 필드별 가중치 검색 (Phase 1.2-B UNION 분해).
|
|
|
|
Phase 1.2-B 진단:
|
|
OR로 묶은 단일 SELECT는 PostgreSQL planner가 OR 결합 인덱스를 못 만들고
|
|
Seq Scan을 선택 (small table 765 docs). EXPLAIN으로 측정 시 525ms.
|
|
→ CTE + UNION으로 분해하면 각 branch가 자기 인덱스 활용 → 26ms (95% 감소).
|
|
|
|
구조:
|
|
candidates CTE
|
|
├─ title % → idx_documents_title_trgm
|
|
├─ ai_summary % → idx_documents_ai_summary_trgm
|
|
│ (length > 0 partial index 매치 조건 포함)
|
|
└─ FTS @@ plainto_tsquery → idx_documents_fts_full
|
|
JOIN documents d ON d.id = c.id
|
|
ORDER BY 5컬럼 similarity 가중 합산 + ts_rank * 2.0
|
|
가중치: title 3.0 / ai_tags 2.5 / user_note 2.0 / ai_summary 1.5 / extracted_text 1.0
|
|
|
|
threshold:
|
|
pg_trgm.similarity_threshold default = 0.3
|
|
→ multi-token 한국어 뉴스 쿼리(예: "이란 미국 전쟁 글로벌 반응")에서
|
|
candidates를 못 모음 → recall 감소 (0.788 → 0.750)
|
|
→ set_limit(0.15)으로 낮춰 recall 회복. precision은 ORDER BY similarity 합산이 보정.
|
|
"""
|
|
from api.search import SearchResult # 순환 import 회피
|
|
|
|
# trigram threshold를 0.15로 낮춰 multi-token query recall 회복
|
|
# SQLAlchemy async session 내 두 execute는 같은 connection 사용
|
|
await session.execute(text("SELECT set_limit(0.15)"))
|
|
|
|
result = await session.execute(
|
|
text("""
|
|
WITH candidates AS (
|
|
-- title trigram (idx_documents_title_trgm)
|
|
SELECT id FROM documents
|
|
WHERE deleted_at IS NULL AND title % :q
|
|
UNION
|
|
-- ai_summary trigram (idx_documents_ai_summary_trgm 부분 인덱스 매치)
|
|
SELECT id FROM documents
|
|
WHERE deleted_at IS NULL
|
|
AND ai_summary IS NOT NULL
|
|
AND length(ai_summary) > 0
|
|
AND ai_summary % :q
|
|
UNION
|
|
-- FTS 통합 인덱스 (idx_documents_fts_full)
|
|
SELECT id FROM documents
|
|
WHERE deleted_at IS NULL
|
|
AND to_tsvector('simple',
|
|
coalesce(title, '') || ' ' ||
|
|
coalesce(ai_tags::text, '') || ' ' ||
|
|
coalesce(ai_summary, '') || ' ' ||
|
|
coalesce(user_note, '') || ' ' ||
|
|
coalesce(extracted_text, '')
|
|
) @@ plainto_tsquery('simple', :q)
|
|
)
|
|
SELECT d.id, d.title, d.ai_domain, d.ai_summary, d.file_format,
|
|
left(d.extracted_text, 200) AS snippet,
|
|
(
|
|
-- 컬럼별 trigram similarity 가중 합산
|
|
similarity(coalesce(d.title, ''), :q) * 3.0
|
|
+ similarity(coalesce(d.ai_tags::text, ''), :q) * 2.5
|
|
+ similarity(coalesce(d.user_note, ''), :q) * 2.0
|
|
+ similarity(coalesce(d.ai_summary, ''), :q) * 1.5
|
|
+ similarity(coalesce(d.extracted_text, ''), :q) * 1.0
|
|
-- FTS 보너스 (idx_documents_fts_full 활용)
|
|
+ coalesce(ts_rank(
|
|
to_tsvector('simple',
|
|
coalesce(d.title, '') || ' ' ||
|
|
coalesce(d.ai_tags::text, '') || ' ' ||
|
|
coalesce(d.ai_summary, '') || ' ' ||
|
|
coalesce(d.user_note, '') || ' ' ||
|
|
coalesce(d.extracted_text, '')
|
|
),
|
|
plainto_tsquery('simple', :q)
|
|
), 0) * 2.0
|
|
) AS score,
|
|
-- match_reason: similarity 가장 큰 컬럼 또는 FTS
|
|
CASE
|
|
WHEN similarity(coalesce(d.title, ''), :q) >= 0.3 THEN 'title'
|
|
WHEN similarity(coalesce(d.ai_tags::text, ''), :q) >= 0.3 THEN 'tags'
|
|
WHEN similarity(coalesce(d.user_note, ''), :q) >= 0.3 THEN 'note'
|
|
WHEN similarity(coalesce(d.ai_summary, ''), :q) >= 0.3 THEN 'summary'
|
|
WHEN similarity(coalesce(d.extracted_text, ''), :q) >= 0.3 THEN 'content'
|
|
ELSE 'fts'
|
|
END AS match_reason
|
|
FROM documents d
|
|
JOIN candidates c ON d.id = c.id
|
|
ORDER BY score DESC
|
|
LIMIT :limit
|
|
"""),
|
|
{"q": query, "limit": limit},
|
|
)
|
|
return [SearchResult(**row._mapping) for row in result]
|
|
|
|
|
|
async def search_vector(
|
|
session: AsyncSession, query: str, limit: int
|
|
) -> list["SearchResult"]:
|
|
"""벡터 유사도 검색 (코사인 거리).
|
|
|
|
Phase 1.2에서 document_chunks 테이블 기반으로 전환 예정.
|
|
현재는 documents.embedding 사용.
|
|
"""
|
|
from api.search import SearchResult # 순환 import 회피
|
|
|
|
try:
|
|
client = AIClient()
|
|
query_embedding = await client.embed(query)
|
|
await client.close()
|
|
except Exception:
|
|
return []
|
|
|
|
result = await session.execute(
|
|
text("""
|
|
SELECT id, title, ai_domain, ai_summary, file_format,
|
|
(1 - (embedding <=> cast(:embedding AS vector))) AS score,
|
|
left(extracted_text, 200) AS snippet,
|
|
'vector' AS match_reason
|
|
FROM documents
|
|
WHERE embedding IS NOT NULL AND deleted_at IS NULL
|
|
ORDER BY embedding <=> cast(:embedding AS vector)
|
|
LIMIT :limit
|
|
"""),
|
|
{"embedding": str(query_embedding), "limit": limit},
|
|
)
|
|
return [SearchResult(**row._mapping) for row in result]
|