altair823
3cd5117a7e
First VectorStore implementation. Per-model Lance tables under config.storage.vector_dir, two-phase upsert (SQLite-pending → Lance MergeInsert → SQLite-committed) with crash-safe retry, search via cosine distance with the spec's score-shift (preserves negative similarity ranking signal that clamping would crush). V003 migration: - Adds status (CHECK constraint pending|committed|tombstone, default pending) and vector_committed columns to embedding_records. - BEFORE DELETE trigger on chunks flips dependent rows to tombstone. Currently overshadowed by V001's ON DELETE CASCADE FK; trigger UPDATE runs first then row vanishes via CASCADE. Spec-faithful tombstone preservation requires recreating embedding_records to drop the CASCADE — deferred to a P+ migration since no production rows exist yet (P3-3 is the first writer). V003 SQL comment explains. LanceVectorStore: - ensure_table is idempotent: opens existing or creates with the Arrow schema (chunk_id, doc_id, embedding FixedSizeList<Float32, dim>, model_id, embedding_version, text, heading_path, created_at). - IndexId computed via id_for_index with collection="chunk_embeddings", index_kind="flat", params_hash = blake3(descriptor JSON). Schema bumps automatically rotate the IndexId. - upsert: phase-1 INSERT OR REPLACE INTO embedding_records (status= 'pending') in a single SQLite tx; phase-2 Lance MergeInsert keyed on chunk_id (idempotent re-run); phase-3 UPDATE status='committed', vector_committed=1. If phase-2 fails the rows stay 'pending' and the next upsert call retries idempotently. - search joins embedding_records WHERE status='committed' so partial- write rows never surface. Cosine distance from Lance ∈ [0, 2] → similarity = 1 - distance ∈ [-1, 1] → score = (similarity + 1)/2 ∈ [0, 1]. NaN coerced to 0 with tracing::warn. Filter by SearchFilters via SqliteStore::filter_chunks (added in this commit). - Sync trait + async LanceDB bridged by an embedded current-thread tokio runtime. Doc-comment on the struct flags the "do NOT call from inside another tokio runtime" panic (block_on cannot nest). kb-app's job scheduler is sync today. kb-store-sqlite additions: - pub fn put_embedding_records_pending(&[EmbeddingRecordRow]) — phase-1 INSERT OR REPLACE (status='pending', vector_committed=0). - pub fn mark_embedding_records_committed(&[EmbeddingId]) — phase-3 single UPDATE … WHERE embedding_id IN (?, ?, …) via params_from_iter, guarded by WHERE status='pending' so tombstones don't get clobbered. - pub fn filter_chunks(&[ChunkId], &SearchFilters) → Vec<ChunkId> consolidates the JOIN against documents/document_tags/ embedding_records + path_glob via globset. Lets kb-store-vector honor SearchFilters without depending on rusqlite or globset directly. (kb-search's filter logic is structurally different — interleaved with the FTS5 SELECT — so it stays as-is for now; consolidation is a P+ refactor.) - 4 new unit tests cover the phase-1 round-trip, empty batch, replay reset of pending rows, and the WHERE-status-pending guard. Tests: - 9 lib unit tests in kb-store-vector covering paths/sanitization, arrow_batch dim validation + descriptor hash, bm25-style cosine score shift math. - 4 new kb-store-sqlite unit tests on filter_chunks (committed-only, tags/lang/trust/path_glob, order preservation, empty input). - 4 new kb-store-sqlite unit tests on the embedding_records helpers. - 8 integration tests in upsert_search.rs and 1 snapshot test marked #[ignore = "requires AVX-capable hardware (LanceDB)"]. They invoke require_avx_or_panic() at the top of each body so a missing-AVX --ignored run fails loudly instead of silently passing. This dev host (qemu64 model) lacks AVX so these were NOT exercised end-to- end here — first CI lane on AVX hardware will validate them. - Snapshot fixture tests/fixtures/vector/run-1.json is a placeholder with an _comment marker. Snapshot test panics until the placeholder is replaced via KB_UPDATE_SNAPSHOTS=1 on AVX hardware. - Workspace 241 passed, 19 ignored, 0 failed; cargo clippy --workspace --all-targets -- -D warnings clean. Allowed deps respected (kb-core, kb-config, kb-store-sqlite, lancedb, arrow + arrow-array + arrow-schema, serde, serde_json, tracing, thiserror) plus forced waivers — anyhow (trait return type), tokio + futures (LanceDB async-only API), blake3 (params_hash). rusqlite and globset are NOT direct deps of kb-store-vector — confirmed via cargo metadata --no-deps. rusqlite stays in [dev-dependencies] for the test fixture seeder only. Out of scope: IVF/PQ index tuning (P+), image vectors (P6), kb-app embed_index orchestration (P3-4 facade). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
kb — Local-first Knowledge Base
상태: spec 동결 단계. 코드 0줄. 30 component task spec 모두 작성/리뷰/머지 완료. 다음 단계 = P0 구현.
kb 는 개인용 로컬 knowledge base + RAG 도구다. Markdown / PDF / 이미지 / 음성을 한 곳에 색인하고, 의미 검색 + citation 포함 LLM 답변을 단일 binary 로 제공한다. 모든 추론은 로컬 (Ollama / fastembed / whisper.cpp) 에서 돌아간다.
대상 하드웨어: M4 48GB MacBook 1대, 사용자 1명.
무엇인가
| 명령 | 동작 |
|---|---|
kb init |
XDG 경로에 데이터 디렉토리 + config.toml 생성 |
kb ingest <path> |
Markdown/PDF/이미지/음성 색인 (idempotent) |
kb search "<query>" |
hybrid (lexical + vector) top-k 검색 — citation 포함 |
kb ask "<query>" |
RAG 답변 + 근거 인용. 근거 부족 시 거절 |
kb inspect doc/chunk <id> |
디버그용 raw record 보기 |
kb doctor |
설정/모델/DB 헬스 체크 |
kb eval run / compare |
golden query 회귀 측정 (chunker/모델 교체 평가) |
기계 친화 모드: 모든 명령에 --json 플래그. 출력은 frozen wire schema v1 (schema_version 필드 항상 포함).
핵심 결정 (lock 됨)
| 결정 | 값 |
|---|---|
| 언어 | Rust 2024 (resolver=3, edition 2024) |
| repo | Cargo workspace (single repo, 함수 호출 기반 모듈러 모놀리스) |
| 원본 저장 | filesystem + blake3 content-addressable copy (대용량은 reference + checksum) |
| metadata | SQLite + FTS5 (lexical search) |
| vector | LanceDB (embedded, model 별 분리 table) |
| Markdown parser | pulldown-cmark |
| embedding | fastembed-rs (multilingual-e5-small, 384d) |
| LLM | Ollama HTTP (default qwen2.5:14b-instruct) |
| 음성 ASR | whisper.cpp (via whisper-rs) |
| OCR | Tesseract (default) + macOS Apple Vision sidecar (feature gate) |
| TUI | Ratatui + crossterm |
| Desktop | Tauri 2 + pdfjs-dist (native PDF render backend 금지) |
| citation 형식 | URI fragment (path#L12-L34, W3C Media Fragments) |
| ID 생성 | blake3(canonical_json(tuple))[..32] hex |
| layout | XDG (~/.local/share/kb/, ~/.config/kb/, …) |
전체는 docs/superpowers/specs/2026-04-27-kb-final-form-design.md 참조.
의존성 그래프
kb-cli, kb-tui, kb-desktop
└─> kb-app
├─> kb-source-fs
├─> kb-parse-md / kb-parse-pdf / kb-parse-image / kb-parse-audio
│ └─> kb-parse-types
├─> kb-normalize
│ └─> kb-parse-types
├─> kb-chunk
├─> kb-store-sqlite
├─> kb-store-vector
├─> kb-embed-local
├─> kb-search
├─> kb-llm-local
├─> kb-rag
├─> kb-eval
└─> kb-config
└─> kb-core (모두 의존)
UI → store/llm/parse 직접 의존 금지. 모든 user-facing 진입은 kb-app facade 만 통한다 (design §8).
Phase 로드맵
| Phase | 내용 | 핵심 산출 crate | 선행 |
|---|---|---|---|
| P0 | Workspace 뼈대 + 도메인 계약 + ID recipe | kb-core, kb-parse-types, kb-config, kb-app, kb-cli |
– |
| P1 | Markdown ingestion (walk → parse → chunk → SQLite) | kb-source-fs, kb-parse-md, kb-normalize, kb-chunk, kb-store-sqlite |
P0 |
| P2 | SQLite FTS5 lexical 검색 + citation | kb-search (lexical) |
P1 |
| P3 | Local embedding + LanceDB + hybrid (RRF) | kb-embed, kb-embed-local, kb-store-vector, kb-search |
P2 |
| P4 | Local LLM + RAG + grounded answer | kb-llm, kb-llm-local, kb-rag |
P3 |
| P5 | Golden query / regression eval | kb-eval |
P4 |
| P6 | 이미지 ingestion (OCR + caption) | kb-parse-image |
P5 |
| P7 | PDF text + page citation | kb-parse-pdf |
P5 |
| P8 | 음성 transcription + timestamp citation | kb-parse-audio |
P5 |
| P9 | TUI + desktop app | kb-tui, kb-desktop |
P5 |
P0P5 직렬. P6P9 P5 이후 병렬 가능.
각 phase 는 component-level 단위로 더 분해되어 있다 (총 30 component task). 자세한 분해는 tasks/INDEX.md.
디렉토리 구조
kb/
├── README.md # 이 파일
├── kb_local_rust_report.md # 최초 설계 보고서 (방향성 + 근거)
├── docs/
│ ├── superpowers/
│ │ ├── specs/
│ │ │ └── 2026-04-27-kb-final-form-design.md # frozen design (12 sections)
│ │ └── plans/
│ │ └── 2026-04-27-task-decomposition.md # task 분해 implementation plan
│ ├── spec/ # P0 에서 생성 — 도메인 모델 / ID / module-boundary 등 stub
│ └── wire-schema/v1/ # P0 에서 생성 — JSON Schema 7개 (citation, search_hit, answer, …)
├── tasks/
│ ├── INDEX.md # phase 인덱스 + component task 트리
│ ├── _template.md # task spec 작성 템플릿
│ ├── phase-0-skeleton.md … phase-9-ui.md # phase epic (high-level)
│ ├── p0/p0-1-skeleton.md # component task (1 spec)
│ ├── p1/p1-1 … p1-6 # component tasks (6)
│ ├── p2/p2-1, p2-2 # (2)
│ ├── p3/p3-1 … p3-4 # (4)
│ ├── p4/p4-1 … p4-3 # (3)
│ ├── p5/p5-1, p5-2 # (2)
│ ├── p6/p6-1 … p6-3 # (3)
│ ├── p7/p7-1, p7-2 # (2)
│ ├── p8/p8-1, p8-2 # (2)
│ └── p9/p9-1 … p9-5 # (5)
├── crates/ # P0 에서 생성 — Rust crates
│ ├── kb-core/
│ ├── kb-parse-types/
│ ├── kb-config/
│ ├── kb-app/
│ └── kb-cli/
└── fixtures/ # P0 에서 생성 — 테스트 fixture 트리
├── markdown/ source-fs/ search/ embed/ vector/
├── rag/ eval/ image/ pdf/ audio/
└── …
빌드 + 실행 (P0 완료 후)
# build
cargo build --release
# 첫 실행
./target/release/kb init
# config 손보고
${EDITOR:-vi} ~/.config/kb/config.toml
# 색인
./target/release/kb ingest ~/KnowledgeBase
# 검색
./target/release/kb search "Markdown chunking 규칙"
# 질문
./target/release/kb ask "내 KB 설계에서 저장소 전략은?"
현재는 P0 미시작 — 위 명령 모두 동작하지 않는다. spec 만 동결됐다.
비-목표 (frozen design §11 / §0)
- 다중 사용자 SaaS, K8s 배포, 원격 vector DB
- enterprise RBAC/ABAC, 실시간 협업
- 모든 파일 포맷의 완벽한 parsing
- agent 가 임의로 파일을 수정하는 자동화
- multi-workspace (P+ 후순위)
- LLM-as-judge eval (rule-based
must_contain만) - visual embedding (CLIP) — P+
- desktop app
kb://protocol handler — P+
외부 AI 통합
kb 의 --json 모드 + frozen wire schema v1 은 외부 자동화의 stable contract. 가능한 통합:
- Claude Code / Codex skill — 얇은 wrapper (
kb search --json/kb ask --json호출). ~50 lines. - MCP server —
kb-mcpbinary (stdio JSON-RPC) 가kb-appfacade 를 1:1 노출. Claude Desktop / Cursor / Zed 등 공유. - HTTP wrapper —
kb serve --bind 127.0.0.1:7711(P+, local-only 가치 깨므로 신중).
기여 / 작업 흐름
이 repo 는 단일 사용자 프로젝트지만 spec 변경 절차는 명문화되어 있다.
- frozen design 변경 —
docs/superpowers/specs/2026-04-27-kb-final-form-design.md가 단일 contract. 변경 시 영향 받는 component task 모두 동시 갱신 필요. PR 1개로 묶기. - 새 component task 추가 —
tasks/_template.md복사 후tasks/p<phase>/p<phase>-<n>-<name>.md생성.contract_sections에 design doc 섹션 명시.Allowed/Forbidden dependencies는 design §8 module-boundary 표 따름. - 구현 — component task 1개당 sub-agent 1세션 권장.
cargo test -p <crate>+ DoD 체크리스트 통과. PR 으로 머지. - 버전 변경 —
parser_version/chunker_version/embedding_version등 변경은 design §9 의 cascade rule 따름. 영향 받는 record 는 재처리 필요.
라이선스
미정 (frozen design 에는 MIT OR Apache-2.0 가 workspace.package 의 license 필드로 권장됨; P0 lock 시 결정).
참고
- 최초 설계 보고서: kb_local_rust_report.md
- Frozen design: docs/superpowers/specs/2026-04-27-kb-final-form-design.md
- Task 분해 plan: docs/superpowers/plans/2026-04-27-task-decomposition.md
- Task 인덱스: tasks/INDEX.md