feat(search): P89 — vector snippet enrich + 관찰 세션 랭킹 강등 (#100)

[hang-in]

Summary

Issue #100 (검색 품질 스모크 테스트) 의 두 항목 fix.

A. vector snippet enrich

• bm25::extract_snippet → pub(crate)
• vector.rs ANN/BLOB 두 경로가 db.get_turn 으로 turn content 가져와 snippet 채움 (이전 String::new()). --vec 결과 snippet: "" 검증성 문제 해소.

B2. 관찰/요약 세션 랭킹 강등

• bm25::SessionMeta 에 turn_count 추가
• reciprocal_rank_fusion 이 turn_count < 3 세션 RRF score *0.5 (제외 아닌 soft 강등). recall 의 automated 완전 제외와 별개 레이어.

Test plan

• cargo fmt --check / clippy --workspace --all-targets -D warnings clean
• cargo test --workspace all green (신규 test_rrf_observer_penalty 포함)

Out of scope (별도)

• hybrid hit 근거 표시 (검색 품질 개선: vector 결과 snippet 공백 및 observer 노이즈 완화 #100 제안 4), 품질 회귀 고정 쿼리 세트 (제안 5)

Closes #100

🤖 Generated with Claude Code

[gemini-code-assist]

Code Review

This pull request introduces search quality improvements by adding a turn_count field to session metadata to demote short observer/summary sessions (less than 3 turns) in the Reciprocal Rank Fusion (RRF) ranking, and by enriching vector search results with snippets extracted from turn content. Feedback on these changes highlights a potential risk where turn_count could become stale during session re-indexing if not synchronized in insert_session_from_vault. Additionally, it is recommended to pass the actual query instead of a hardcoded empty string to extract_snippet in vector search to enable context-aware snippet highlighting.

[gemini-code-assist]

Stale turn_count Risk on Re-indexing:
The turn_count retrieved here can become stale for sessions that are re-indexed from the vault. In insert_session_from_vault, an INSERT OR IGNORE statement is used, and the subsequent UPDATE only synchronizes is_archived and archived_at. If a session's turn count changes in the vault and is re-indexed, the turn_count in the database will not be updated, leading to incorrect ranking demotions in RRF.

Consider updating turn_count (along with other metadata like summary) during the re-indexing update phase in insert_session_from_vault.

[gemini-code-assist]

Sub-optimal Snippet Extraction (Hardcoded Empty Query):
Passing an empty string "" as the query to extract_snippet means the snippet will always be extracted from the very beginning of the turn content (since find("") returns Some(0)). If the user's query matches terms later in the turn content, those matching terms won't be highlighted or centered in the snippet, which degrades the snippet quality for vector search results.

While changing the signature of search_with_embedding to accept an optional query string query: Option<&str> is outside the scope of the current diff hunks, consider implementing this as a future improvement so that the actual query can be passed down from SearchEngine::search to provide context-aware snippets.

[hang-in]

Gemini 리뷰 2건 (N+1) 반영 (commit 51256e5). get_turn_contents 배치 메서드 (row-value IN (VALUES ...) 단일 쿼리) 추가 + fill_snippets helper 로 vector ANN/BLOB 두 경로의 DB 왕복을 N→1 로. 신규 unit test 2건. fmt + clippy + 전체 test clean.

[hang-in]

Gemini 리뷰 반영:

• ✅ Feature: Add summary field to session frontmatter for better Dataview discoverability / 세션 frontmatter에 summary 필드 추가 제안 #1 (turn_count stale): insert_session_from_vault 의 reindex UPDATE 에 turn_count 동기화 추가 (commit f0a3c47) + 회귀 test.
• ⏭️ Escape :: patterns in session body to prevent Dataview inline field pollution #2 (query-aware snippet): search_with_embedding 은 의도적으로 query 문자열 없는 pre-computed embedding 경로라, query highlight 를 위해 hybrid/server 까지 시그니처를 확장하는 별도 작업이 필요합니다. 현재는 vector search() (async, query 보유) 경로가 이미 적용돼 있고, pre-computed 경로는 turn 앞부분 snippet 으로 충분히 검증성을 확보합니다. follow-up 으로 분리합니다.