feat(embed): candle 임베딩 provider (NUMA-안전, opt-in) + v0.22.0

duo-socket NUMA 서버에서 fastembed(onnxruntime)가 intra-op 스레드를 48개로
하드코딩해 NUMA 힙 손상 → double-free 로 ingest 가 죽는 문제를 회피하기 위해,
같은 multilingual-e5-large 모델을 순수 Rust(candle)로 돌리는 opt-in 임베딩
provider 를 추가한다.

- 신규 crate kebab-embed-candle: CandleEmbedder (kebab_core::Embedder).
  hf-hub safetensors → XLMRobertaModel forward → mask mean-pool → L2 → e5
  prefix. candle 의존성 트리를 이 crate 에 격리 (core/config 외 kebab-* 의존 0).
- 스레드 캡: [models.embedding].num_threads + env KEBAB_EMBED_THREADS →
  글로벌 rayon 풀 1회 캡 (NUMA-안전 레버).
- kebab-app::embedder() 가 provider 분기 (fastembed/onnx/"" → 기존 경로 불변,
  candle → CandleEmbedder, 미지값 → 에러).
- Phase 0 스파이크 crate 제거 (production 흡수).
- 버전 0.21.1 → 0.22.0 (신규 config surface, pre-1.0 minor bump).

패리티: cosine_min=1.000000, max abs diff=2.01e-7 (< 1e-5) → embedding_version
유지, 재색인 0. fastembed default 동작/벡터 불변. wire schema 변경 없음.

검증(파일+exit code): clippy -D warnings EXIT=0(warning 0), test EXIT=0
(candle unit 5 + thread_cap rayon=4 + config 68), parity #[ignore] EXIT=0.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This commit is contained in:
2026-06-01 14:52:25 +00:00
parent 76841af7d3
commit 8f7b6ee538
18 changed files with 825 additions and 330 deletions

81
Cargo.lock generated
View File

@@ -4530,7 +4530,7 @@ dependencies = [
[[package]]
name = "kebab-app"
version = "0.21.1"
version = "0.22.0"
dependencies = [
"anyhow",
"base64 0.22.1",
@@ -4543,6 +4543,7 @@ dependencies = [
"kebab-config",
"kebab-core",
"kebab-embed",
"kebab-embed-candle",
"kebab-embed-local",
"kebab-llm",
"kebab-llm-local",
@@ -4576,7 +4577,7 @@ dependencies = [
[[package]]
name = "kebab-chunk"
version = "0.21.1"
version = "0.22.0"
dependencies = [
"anyhow",
"blake3",
@@ -4594,7 +4595,7 @@ dependencies = [
[[package]]
name = "kebab-cli"
version = "0.21.1"
version = "0.22.0"
dependencies = [
"anyhow",
"clap",
@@ -4615,7 +4616,7 @@ dependencies = [
[[package]]
name = "kebab-config"
version = "0.21.1"
version = "0.22.0"
dependencies = [
"anyhow",
"dirs 5.0.1",
@@ -4631,7 +4632,7 @@ dependencies = [
[[package]]
name = "kebab-core"
version = "0.21.1"
version = "0.22.0"
dependencies = [
"anyhow",
"blake3",
@@ -4645,7 +4646,7 @@ dependencies = [
[[package]]
name = "kebab-embed"
version = "0.21.1"
version = "0.22.0"
dependencies = [
"anyhow",
"blake3",
@@ -4657,9 +4658,28 @@ dependencies = [
"tracing",
]
[[package]]
name = "kebab-embed-candle"
version = "0.22.0"
dependencies = [
"anyhow",
"candle-core",
"candle-nn",
"candle-transformers",
"hf-hub",
"kebab-config",
"kebab-core",
"kebab-embed-local",
"rayon",
"serde_json",
"tempfile",
"tokenizers 0.21.4",
"tracing",
]
[[package]]
name = "kebab-embed-local"
version = "0.21.1"
version = "0.22.0"
dependencies = [
"anyhow",
"fastembed",
@@ -4672,7 +4692,7 @@ dependencies = [
[[package]]
name = "kebab-eval"
version = "0.21.1"
version = "0.22.0"
dependencies = [
"anyhow",
"kebab-app",
@@ -4691,7 +4711,7 @@ dependencies = [
[[package]]
name = "kebab-llm"
version = "0.21.1"
version = "0.22.0"
dependencies = [
"anyhow",
"kebab-core",
@@ -4700,7 +4720,7 @@ dependencies = [
[[package]]
name = "kebab-llm-local"
version = "0.21.1"
version = "0.22.0"
dependencies = [
"anyhow",
"kebab-config",
@@ -4717,7 +4737,7 @@ dependencies = [
[[package]]
name = "kebab-mcp"
version = "0.21.1"
version = "0.22.0"
dependencies = [
"anyhow",
"kebab-app",
@@ -4735,7 +4755,7 @@ dependencies = [
[[package]]
name = "kebab-nli"
version = "0.21.1"
version = "0.22.0"
dependencies = [
"anyhow",
"hf-hub",
@@ -4750,7 +4770,7 @@ dependencies = [
[[package]]
name = "kebab-parse-code"
version = "0.21.1"
version = "0.22.0"
dependencies = [
"anyhow",
"gix",
@@ -4773,7 +4793,7 @@ dependencies = [
[[package]]
name = "kebab-parse-image"
version = "0.21.1"
version = "0.22.0"
dependencies = [
"ab_glyph",
"anyhow",
@@ -4797,7 +4817,7 @@ dependencies = [
[[package]]
name = "kebab-parse-md"
version = "0.21.1"
version = "0.22.0"
dependencies = [
"anyhow",
"kebab-core",
@@ -4814,7 +4834,7 @@ dependencies = [
[[package]]
name = "kebab-parse-pdf"
version = "0.21.1"
version = "0.22.0"
dependencies = [
"anyhow",
"blake3",
@@ -4829,7 +4849,7 @@ dependencies = [
[[package]]
name = "kebab-rag"
version = "0.21.1"
version = "0.22.0"
dependencies = [
"anyhow",
"blake3",
@@ -4851,7 +4871,7 @@ dependencies = [
[[package]]
name = "kebab-search"
version = "0.21.1"
version = "0.22.0"
dependencies = [
"anyhow",
"globset",
@@ -4870,7 +4890,7 @@ dependencies = [
[[package]]
name = "kebab-source-fs"
version = "0.21.1"
version = "0.22.0"
dependencies = [
"anyhow",
"blake3",
@@ -4888,7 +4908,7 @@ dependencies = [
[[package]]
name = "kebab-store-sqlite"
version = "0.21.1"
version = "0.22.0"
dependencies = [
"anyhow",
"blake3",
@@ -4908,7 +4928,7 @@ dependencies = [
[[package]]
name = "kebab-store-vector"
version = "0.21.1"
version = "0.22.0"
dependencies = [
"anyhow",
"arrow",
@@ -4932,7 +4952,7 @@ dependencies = [
[[package]]
name = "kebab-tui"
version = "0.21.1"
version = "0.22.0"
dependencies = [
"anyhow",
"crossterm",
@@ -8397,23 +8417,6 @@ dependencies = [
"smallvec",
]
[[package]]
name = "spike-embed-candle"
version = "0.0.0"
dependencies = [
"anyhow",
"candle-core",
"candle-nn",
"candle-transformers",
"hf-hub",
"kebab-config",
"kebab-embed",
"kebab-embed-local",
"rayon",
"serde_json",
"tokenizers 0.21.4",
]
[[package]]
name = "spm_precompiled"
version = "0.1.4"

View File

@@ -11,6 +11,7 @@ members = [
"crates/kebab-search",
"crates/kebab-embed",
"crates/kebab-embed-local",
"crates/kebab-embed-candle",
"crates/kebab-llm",
"crates/kebab-llm-local",
"crates/kebab-rag",
@@ -23,8 +24,6 @@ members = [
"crates/kebab-mcp",
"crates/kebab-parse-code",
"crates/kebab-nli",
# Track 1 / Phase 0 feasibility spike (throwaway; candle deps isolated here).
"crates/spike-embed-candle",
]
[workspace.package]
@@ -32,7 +31,7 @@ edition = "2024"
rust-version = "1.85"
license = "MIT OR Apache-2.0"
repository = "https://github.com/altair823/kebab"
version = "0.21.1" # v0.21.1 — config 마이그레이션(kebab config migrate): 기존 config.toml 에 빠진 섹션 주석과 함께 추가 + deprecated 정리 + schema_version 1→2 — CLAUDE.md §Release 도그푸딩 트리거
version = "0.22.0" # v0.22.0 — candle 임베딩 provider (NUMA-안전, opt-in `provider=candle` + `num_threads`/KEBAB_EMBED_THREADS). fastembed default 불변, embedding_version 유지(재색인 0). — CLAUDE.md §Release 도그푸딩 트리거
# pre-v0.18 workspace-wide cleanup: enable clippy::pedantic group with
# intentional allow-list. The allowed lints are either cosmetic (doc style),

View File

@@ -30,6 +30,7 @@ P0~P5 직렬. P6~P9 P5 이후 병렬 가능.
## 머지 후 발견된 버그 / 결정 (요약)
- **candle 임베딩 백엔드 다변화** (2026-06-01, Track 1, v0.22.0): `provider = "candle"` opt-in 추가 — 같은 `multilingual-e5-large` 모델을 순수 Rust(candle)로 돌려 듀얼소켓 NUMA 서버의 onnxruntime 48-스레드 double-free 를 회피. `[models.embedding].num_threads`(+env `KEBAB_EMBED_THREADS`)로 CPU 스레드 캡. fastembed default 동작·벡터 불변, `embedding_version` 유지(재색인 0). Phase 0 스파이크 패리티 cosine 1.000000. 상세 HOTFIXES 동일 일자.
- **config 마이그레이션** (2026-05-31, PR #198): `kebab config migrate` 추가 — 기존 config.toml 에 빠진 섹션을 주석과 함께 채우고 deprecated 정리(멱등·`.bak`·dry-run, 값/주석 보존). `schema_version` 1→2, `init` 도 섹션 주석 포함, doctor 에 `config_migration` 체크. 상세 HOTFIXES 동일 일자.
머지 후 발견된 모든 deviation / hotfix 의 dated 로그는 [tasks/HOTFIXES.md](tasks/HOTFIXES.md). 본 요약은 \"누군가가 인수받을 때 알아두면 시간을 많이 절약하는\" 항목만:

85
IMPL_REPORT.md Normal file
View File

@@ -0,0 +1,85 @@
# Track 1 / Phase 1 — candle 임베딩 provider 구현 보고서
- 날짜: 2026-06-01
- 브랜치: `feat/embed-candle` (worktree `/build/out/kebab-worktrees/embed-candle`)
- 스펙: `docs/superpowers/specs/2026-06-01-embed-candle-track-spec.md`
- 버전: 0.21.1 → **0.22.0**
## 1. 변경 요약
| 영역 | 변경 |
|------|------|
| 신규 crate | `crates/kebab-embed-candle``CandleEmbedder` (`kebab_core::Embedder` impl). 스파이크 파이프라인 흡수: safetensors via hf-hub → `XLMRobertaModel` forward(`Device::Cpu`) → attention-mask mean pooling → L2 → e5 prefix(`passage:`/`query:`). 모델 캐시 `{model_dir}/candle/`. deps = candle-core/nn/transformers 0.10.2, tokenizers, hf-hub, serde_json, rayon, anyhow, tracing + `kebab-core`/`kebab-config` 만 (design §8 경계 준수). |
| 스레드 캡 | `[models.embedding].num_threads: u32`(default 0=auto) + env `KEBAB_EMBED_THREADS`(우선). `apply_thread_cap()` 가 글로벌 rayon 풀 1회 캡 (이미 init 시 no-op). |
| 주입 분기 | `kebab-app::App::embedder()``provider` 분기 — `fastembed`/`onnx`/`""` → 기존 `FastembedEmbedder`(불변), `candle``CandleEmbedder`, 미지값 → 에러. `none` 은 기존 lexical-only. `kebab-app/Cargo.toml` 에 dep 추가. |
| config | `EmbeddingModelCfg.num_threads`(`#[serde(default)]` — 옛 config 호환) + `KEBAB_MODELS_EMBEDDING_NUM_THREADS` env + `Config::defaults()`. |
| 스파이크 제거 | `crates/spike-embed-candle` 삭제 + 워크스페이스 멤버 제거 + `spike_build.log`/`spike_run.log` 정리. |
| 문서/버전 | README Configuration, `docs/SMOKE.md` config 예시, `docs/ARCHITECTURE.md`(crate 그래프+트리), HANDOFF 한 줄, `tasks/HOTFIXES.md` 2026-06-01 dated entry, workspace `version` 0.22.0, `docs/release-notes/v0.22.0-draft.md`. |
## 2. 검증 게이트 결과 (모두 파일 출력 + exit code 로 검증)
> ⚠️ 주의: background shell 의 notification "exit 0" 은 wrapper 의 종료코드라
> 신뢰 불가. 실제 결과는 각 로그의 `*_EXIT=` 라인 값으로 확정했다
> ([[project_rerank_experiment]] 교훈). 실제로 첫 빌드는 wrapper 가 exit 0 을
> 보고했지만 로그의 `BUILD_EXIT=101`(serde_json 미선언)이었고, dep 추가 후 통과.
| 게이트 | 명령 | 결과 | 로그 |
|--------|------|------|------|
| 빌드 + clippy | `cargo clippy --workspace --all-targets -j 4 -- -D warnings` | **`CLIPPY_EXIT=0`**, warning 0 | `clippy.log` |
| 단위/통합 테스트 | `cargo test -p kebab-embed-candle -p kebab-config -j 4` | **`TEST_EXIT=0`** — candle lib unit 5, `thread_cap` 1 passed(rayon current=4 검증), config 68 passed, parity 1 ignored | `test_units.log` |
| config 회귀 | (위 동일 run, `kebab-config` 68 tests) | 0 failed | `test_units.log` |
| 패리티 `#[ignore]` 수동 1회 | `cargo test -p kebab-embed-candle --release -- --ignored --nocapture` | **`PARITY_EXIT=0`**, 1 passed (32.53s) | `test_parity.log` |
## 3. 패리티 수치 (재색인 결정 근거 — 스펙 D-reindex)
10 문장(한/영 혼합) candle vs `FastembedEmbedder`(onnxruntime):
```
PARITY_SUMMARY cosine_min=1.000000 max_abs_diff=2.011657e-7
```
- 코사인 최소 **1.000000** (≥ 0.9999 게이트 통과).
- 차원별 **max 절대오차 = 2.01e-7** — 스펙이 정한 "사실상 동일" 기준
(max abs diff < 1e-5) 보다 **약 50배 작다**.
- **결론: `embedding_version` 유지 = 재색인 0.** candle 과 onnxruntime 의
벡터는 f32 반올림 수준에서만 다르며 (e-7), 기존 LanceDB 색인을 그대로
재사용해도 검색 결과가 바뀌지 않는다. version bump / cascade 트리거 안 함.
## 4. 잔여 리스크
- **CPU latency**: candle 는 순수 Rust 라 onnxruntime 의 네이티브 커널보다
느리다 (Phase 0 스파이크 ~4×). 그래서 default 는 fastembed 유지, candle 은
NUMA 환경 opt-in. 단일 워크스테이션 사용자에게는 권하지 않음 (README 명시).
- **모델 다운로드**: candle 은 `{model_dir}/candle/` 에 safetensors(~2GB)를
별도 캐시 (onnx 캐시와 공유 안 함). 첫 ingest 시 ~2GB 다운로드 발생.
- **잔여 게이트 (사용자 실행, Claude 불가, meta-spec §4.3)**: 그 듀얼소켓
NUMA 서버에서 `provider=candle` 로 5150-doc ingest 가 double-free 없이
EXIT=0 완주하는지 — 이 머신은 GPU/NUMA 없는 단일 VM 이라 재현 불가. PR
머지 전/후 사용자 검증 예약.
- **골든 스위트 회귀 0 (스펙 §8)**: provider=candle 로 `kebab-eval` 골든
스위트 실행은 본 worktree 범위 밖(사용자 도그푸딩 단계). 패리티 e-7 로
벡터 동일성이 입증돼 회귀 위험은 낮음.
## 5. 재현 명령
```bash
cd /build/out/kebab-worktrees/embed-candle
export CARGO_TARGET_DIR=/build/out/cargo-target/target
# 빌드 + clippy (warning 0)
cargo clippy --workspace --all-targets -j 4 -- -D warnings
# 단위 + config 회귀
cargo test -p kebab-embed-candle -p kebab-config -j 4
# 패리티 (모델 ~2GB 다운로드 + 양쪽 추론, /build/dogfood/config.toml 필요)
cargo test -p kebab-embed-candle --release -j 4 -- --ignored --nocapture
# → PARITY_SUMMARY cosine_min=1.000000 max_abs_diff=2.011657e-7
# candle provider 로 ingest (사용자 NUMA 검증)
KEBAB_EMBED_THREADS=8 kebab ingest --config /path/to/candle-config.toml
```
## 6. 커밋
`feat/embed-candle` 에 커밋 완료. push / PR 은 메인 세션이 처리 (본 worker 는 하지 않음).

View File

@@ -97,9 +97,17 @@ root = "~/KnowledgeBase" # 색인할 폴더. 절대 / tilde / env / 상대 경
# 상대 경로의 base 는 config.toml 위치 (cwd 무관).
[models.embedding]
provider = "fastembed" # "fastembed"(기본, onnxruntime) / "candle"(순수 Rust)
# / "none"(lexical-only). candle 는 같은 모델·같은 벡터를
# 순수 Rust 로 돌려 NUMA 서버의 onnxruntime 48-스레드
# double-free 를 피하는 opt-in 백엔드 (재색인 불필요).
model = "multilingual-e5-large" # 다국어 sentence embedding (1024-dim).
# 첫 ingest 시 ONNX (~1.3GB) 자동 다운로드.
# candle provider 는 safetensors (~2GB) 다운로드.
dimensions = 1024 # config 와 LanceDB stored dim 불일치 시 검색 0건.
num_threads = 0 # candle 전용 CPU 스레드 캡 (0=auto=#cores).
# env KEBAB_EMBED_THREADS 가 우선. NUMA 노드 바인딩은
# numactl 과 조합. fastembed provider 는 무시.
[models.llm]
endpoint = "http://localhost:11434" # Ollama host:port

View File

@@ -18,6 +18,7 @@ kebab-store-vector = { path = "../kebab-store-vector" }
kebab-search = { path = "../kebab-search" }
kebab-embed = { path = "../kebab-embed" }
kebab-embed-local = { path = "../kebab-embed-local" }
kebab-embed-candle = { path = "../kebab-embed-candle" }
kebab-llm = { path = "../kebab-llm" }
kebab-llm-local = { path = "../kebab-llm-local" }
kebab-rag = { path = "../kebab-rag" }

View File

@@ -43,6 +43,7 @@ use kebab_core::{
Answer, DocumentStore, Embedder, ExtractContext, Extractor, IndexVersion, LanguageModel,
MediaType, Retriever, SearchHit, SearchMode, SearchOpts, SearchQuery, VectorStore,
};
use kebab_embed_candle::CandleEmbedder;
use kebab_embed_local::FastembedEmbedder;
use kebab_llm_local::OllamaLanguageModel;
use kebab_parse_code::{
@@ -833,9 +834,26 @@ impl App {
if let Some(e) = self.embedder.get() {
return Ok(Some(e.clone()));
}
let emb: Arc<dyn Embedder + Send + Sync> = Arc::new(
FastembedEmbedder::new(&self.config).context("kb-app: load FastembedEmbedder")?,
);
// Provider branch (Track 1 spec §3). `embeddings_disabled()` above
// already handled `"none"`; here we route the live providers.
// `fastembed`/`onnx`/(empty) keep the default onnxruntime path
// (vectors unchanged — `embedding_version` is preserved); `candle`
// selects the pure-Rust NUMA-safe backend.
let provider = self.config.models.embedding.provider.as_str();
let emb: Arc<dyn Embedder + Send + Sync> = match provider {
"fastembed" | "onnx" | "" => Arc::new(
FastembedEmbedder::new(&self.config).context("kb-app: load FastembedEmbedder")?,
),
"candle" => Arc::new(
CandleEmbedder::new(&self.config).context("kb-app: load CandleEmbedder")?,
),
other => {
return Err(anyhow!(
"kb-app: unknown embedding provider {other:?}; expected one of \
`fastembed` (default), `candle`, or `none` (lexical-only)"
));
}
};
// `set` returns Err if another thread won the race; in that case
// the loser still returns the (now-cached) winner via `get()`.
let _ = self.embedder.set(emb.clone());

View File

@@ -155,11 +155,21 @@ impl NliCfg {
#[derive(Clone, Debug, PartialEq, Serialize, Deserialize)]
pub struct EmbeddingModelCfg {
/// `fastembed` (default, onnxruntime) or `candle` (pure-Rust,
/// NUMA-safe). `none` disables embeddings (lexical-only). Unknown
/// values error at embedder construction.
pub provider: String,
pub model: String,
pub version: String,
pub dimensions: usize,
pub batch_size: usize,
/// Cap on the CPU worker threads the `candle` provider spins up
/// (sizes the global rayon pool; env `KEBAB_EMBED_THREADS` overrides).
/// `0` = auto (rayon default = #cores). Lever to sidestep the
/// onnxruntime 48-thread NUMA double-free; ignored by the `fastembed`
/// provider. Defaulted on load so pre-0.22 config files still parse.
#[serde(default)]
pub num_threads: u32,
}
#[derive(Clone, Debug, PartialEq, Serialize, Deserialize)]
@@ -707,6 +717,7 @@ impl Config {
version: "v1".to_string(),
dimensions: 1024,
batch_size: 64,
num_threads: 0,
},
llm: LlmCfg {
provider: "ollama".to_string(),
@@ -964,6 +975,11 @@ impl Config {
self.models.embedding.batch_size = n;
}
}
"KEBAB_MODELS_EMBEDDING_NUM_THREADS" => {
if let Ok(n) = v.parse::<u32>() {
self.models.embedding.num_threads = n;
}
}
// models.llm
"KEBAB_MODELS_LLM_PROVIDER" => self.models.llm.provider = v.clone(),

View File

@@ -0,0 +1,39 @@
[package]
name = "kebab-embed-candle"
version = { workspace = true }
edition = { workspace = true }
rust-version = { workspace = true }
license = { workspace = true }
repository = { workspace = true }
description = "Pure-Rust candle adapter implementing kb_core::Embedder (multilingual-e5-large, NUMA-safe thread cap)"
[dependencies]
kebab-core = { path = "../kebab-core" }
kebab-config = { path = "../kebab-config" }
# candle stack — pinned to the workspace-locked crates.io release (0.10.x),
# same versions the Phase 0 spike compiled so build artifacts are reused.
candle-core = "0.10.2"
candle-nn = "0.10.2"
candle-transformers = "0.10.2"
tokenizers = "0.21"
hf-hub = { version = "0.4", features = ["ureq"] }
serde_json = { workspace = true }
# Thread cap: a one-shot global rayon pool sizes candle's CPU threads
# (the Phase 0 spike proved RAYON_NUM_THREADS caps candle), so a NUMA host
# can keep onnxruntime's hard-coded 48-intra-op heap corruption at bay.
rayon = "1"
anyhow = { workspace = true }
tracing = { workspace = true }
[dev-dependencies]
# Integration-test binaries can only see the library's public API + these,
# not the library's own (non-dev) dependencies — so rayon/kebab-config/kebab-core
# are repeated here for tests/parity.rs and tests/thread_cap.rs.
kebab-embed-local = { path = "../kebab-embed-local" }
kebab-config = { path = "../kebab-config" }
kebab-core = { path = "../kebab-core" }
rayon = "1"
tempfile = { workspace = true }
[lints]
workspace = true

View File

@@ -0,0 +1,363 @@
//! `kebab-embed-candle` — [`CandleEmbedder`], a pure-Rust (candle)
//! implementation of [`Embedder`](kebab_core::Embedder).
//!
//! Runs the same `intfloat/multilingual-e5-large` model as the default
//! [`FastembedEmbedder`](kebab_embed_local) but through `candle`
//! (`candle-transformers`' XLM-RoBERTa) instead of onnxruntime. Motivation:
//! fastembed 4.9's onnxruntime hard-codes 48 intra-op threads, which corrupts
//! the heap (double-free) on dual-socket NUMA hosts. candle's CPU backend
//! sizes its threads off the global rayon pool, so a one-shot
//! [`rayon::ThreadPoolBuilder`] cap (config `num_threads` / env
//! `KEBAB_EMBED_THREADS`) keeps the worker count NUMA-safe.
//!
//! Output parity with the onnxruntime path was proven by the Phase 0 spike
//! (cosine 1.000000); this crate absorbs that pipeline verbatim:
//!
//! 1. e5 prefix (`passage: ` for documents, `query: ` for queries — the same
//! convention as `kebab-embed-local`'s `prefix_input`);
//! 2. tokenize (max_len 512, batch-longest padding, special tokens);
//! 3. XLM-RoBERTa forward on `Device::Cpu`;
//! 4. attention-mask-weighted mean pooling;
//! 5. L2 normalization.
//!
//! Model files (`config.json`, `tokenizer.json`, `model.safetensors`) are
//! fetched via `hf-hub` into `{config.storage.model_dir}/candle/`.
//!
//! This crate is **opt-in** (`config.models.embedding.provider = "candle"`);
//! the default provider stays `fastembed`. See
//! `docs/superpowers/specs/2026-06-01-embed-candle-track-spec.md`.
use std::sync::Mutex;
use anyhow::{Context, Result};
use candle_core::{DType, Device, Tensor};
use candle_nn::VarBuilder;
use candle_transformers::models::xlm_roberta::{Config as XlmConfig, XLMRobertaModel};
use kebab_config::{Config, expand_path};
use kebab_core::{Embedder, EmbeddingInput, EmbeddingKind, EmbeddingModelId, EmbeddingVersion};
use tokenizers::{PaddingParams, PaddingStrategy, Tokenizer, TruncationParams};
/// Subdirectory under `config.storage.model_dir` where the candle adapter
/// caches safetensors + tokenizer. Mirrors `kebab-embed-local`'s
/// `fastembed/` subdir so the two backends never collide.
const CANDLE_CACHE_SUBDIR: &str = "candle";
/// HuggingFace repo id for the multilingual e5 large model. Same weights the
/// onnxruntime path uses, just the safetensors variant candle can read.
const HF_MODEL: &str = "intfloat/multilingual-e5-large";
/// Token truncation length (e5 was trained at 512).
const MAX_LEN: usize = 512;
/// Env var that overrides `config.models.embedding.num_threads`. Read once in
/// [`CandleEmbedder::new`]; `0`/unset/unparseable means "leave rayon default".
const ENV_EMBED_THREADS: &str = "KEBAB_EMBED_THREADS";
/// Pure-Rust candle adapter. Construct via [`CandleEmbedder::new`]; the
/// constructor downloads the model on first use, so share one instance.
pub struct CandleEmbedder {
// candle's `forward` is `&self`, but `XLMRobertaModel` is not guaranteed
// `Sync`; the `Mutex` both supplies that bound and serializes inference
// (callers batch sequentially anyway — same rationale as
// `FastembedEmbedder`).
model: Mutex<XLMRobertaModel>,
tokenizer: Tokenizer,
device: Device,
model_id: EmbeddingModelId,
version: EmbeddingVersion,
dimensions: usize,
batch_size: usize,
}
impl CandleEmbedder {
/// Build an embedder from `Config`. Applies the NUMA thread cap, fetches
/// the model into `{model_dir}/candle/`, and validates that the model's
/// hidden size matches `config.models.embedding.dimensions` before
/// returning.
pub fn new(config: &Config) -> Result<Self> {
// 1. NUMA thread cap. env `KEBAB_EMBED_THREADS` wins over the config
// field; `0`/unset leaves rayon's default. `build_global` errors if
// the pool was already initialized — intentionally ignored so a
// second embedder (or a prior rayon user) is a no-op, not a failure.
let n_threads = std::env::var(ENV_EMBED_THREADS)
.ok()
.and_then(|v| v.parse::<usize>().ok())
.unwrap_or(config.models.embedding.num_threads as usize);
if n_threads > 0 {
if apply_thread_cap(n_threads) {
tracing::info!(
target: "kebab-embed-candle",
num_threads = n_threads,
"capped global rayon pool for candle CPU backend"
);
} else {
tracing::debug!(
target: "kebab-embed-candle",
requested = n_threads,
"global rayon pool already initialized; thread cap not applied"
);
}
}
// 2. Resolve `{data_dir}/models/candle/` exactly like the fastembed
// adapter resolves its own subdir.
let data_dir = expand_path(&config.storage.data_dir, "");
let model_dir = expand_path(&config.storage.model_dir, &data_dir.to_string_lossy());
let cache_dir = model_dir.join(CANDLE_CACHE_SUBDIR);
std::fs::create_dir_all(&cache_dir)
.with_context(|| format!("create candle cache dir {}", cache_dir.display()))?;
let device = Device::Cpu;
// 3. Fetch model files via hf-hub into the candle cache.
tracing::info!(
target: "kebab-embed-candle",
cache_dir = %cache_dir.display(),
model = HF_MODEL,
"loading candle embedding model (first run downloads ~2GB safetensors)"
);
let api = hf_hub::api::sync::ApiBuilder::new()
.with_cache_dir(cache_dir.clone())
.build()
.context("kb-embed-candle: build hf-hub api")?;
let repo = api.model(HF_MODEL.to_string());
let config_path = repo.get("config.json").context("download config.json")?;
let tokenizer_path = repo
.get("tokenizer.json")
.context("download tokenizer.json")?;
let weights_path = repo
.get("model.safetensors")
.context("download model.safetensors")?;
// 4. Build the candle XLM-RoBERTa model.
let cfg_json = std::fs::read_to_string(&config_path)
.with_context(|| format!("read {}", config_path.display()))?;
let cfg: XlmConfig =
serde_json::from_str(&cfg_json).context("kb-embed-candle: parse XLM-R config")?;
// Validate dim BEFORE building the model so a misconfigured
// `dimensions` fails cheaply (matches FastembedEmbedder's contract).
check_dim(cfg.hidden_size, config.models.embedding.dimensions)?;
let vb = unsafe {
VarBuilder::from_mmaped_safetensors(&[weights_path], DType::F32, &device)
.context("kb-embed-candle: mmap safetensors")?
};
let model =
XLMRobertaModel::new(&cfg, vb).context("kb-embed-candle: build XLMRobertaModel")?;
let mut tokenizer = Tokenizer::from_file(&tokenizer_path)
.map_err(|e| anyhow::anyhow!("kb-embed-candle: load tokenizer: {e}"))?;
tokenizer
.with_padding(Some(PaddingParams {
strategy: PaddingStrategy::BatchLongest,
..Default::default()
}))
.with_truncation(Some(TruncationParams {
max_length: MAX_LEN,
..Default::default()
}))
.map_err(|e| anyhow::anyhow!("kb-embed-candle: set truncation: {e}"))?;
tracing::info!(
target: "kebab-embed-candle",
dimensions = cfg.hidden_size,
layers = cfg.num_hidden_layers,
"candle embedding model loaded"
);
Ok(Self {
model: Mutex::new(model),
tokenizer,
device,
model_id: EmbeddingModelId(config.models.embedding.model.clone()),
version: EmbeddingVersion(config.models.embedding.version.clone()),
dimensions: cfg.hidden_size,
batch_size: config.models.embedding.batch_size.max(1),
})
}
/// Embed one batch (already prefixed) through the candle pipeline:
/// tokenize → forward → masked mean pool → L2 normalize.
fn embed_batch(&self, prefixed: &[String]) -> Result<Vec<Vec<f32>>> {
let encodings = self
.tokenizer
.encode_batch(prefixed.to_vec(), true)
.map_err(|e| anyhow::anyhow!("kb-embed-candle: encode_batch: {e}"))?;
let bsz = encodings.len();
let seq = encodings[0].get_ids().len();
let mut ids = Vec::with_capacity(bsz * seq);
let mut mask = Vec::with_capacity(bsz * seq);
for enc in &encodings {
ids.extend(enc.get_ids().iter().copied());
mask.extend(enc.get_attention_mask().iter().map(|&m| m as f32));
}
let input_ids = Tensor::from_vec(ids, (bsz, seq), &self.device)?;
let attn_f32 = Tensor::from_vec(mask, (bsz, seq), &self.device)?;
let token_type_ids = input_ids.zeros_like()?;
let hidden = {
let guard = self
.model
.lock()
.unwrap_or_else(std::sync::PoisonError::into_inner);
// forward: (input_ids, attention_mask, token_type_ids, past,
// encoder_hidden, encoder_mask)
guard.forward(&input_ids, &attn_f32, &token_type_ids, None, None, None)?
};
// attention-mask-weighted mean pooling
let mask3 = attn_f32.unsqueeze(2)?; // (b, seq, 1)
let summed = hidden.broadcast_mul(&mask3)?.sum(1)?; // (b, hidden)
let counts = mask3.sum(1)?; // (b, 1)
let mean = summed.broadcast_div(&counts)?;
// L2 normalize
let norm = mean.sqr()?.sum_keepdim(1)?.sqrt()?;
let normalized = mean.broadcast_div(&norm)?;
Ok(normalized.to_vec2::<f32>()?)
}
}
impl Embedder for CandleEmbedder {
fn model_id(&self) -> EmbeddingModelId {
self.model_id.clone()
}
fn model_version(&self) -> EmbeddingVersion {
self.version.clone()
}
fn dimensions(&self) -> usize {
self.dimensions
}
fn embed(&self, inputs: &[EmbeddingInput<'_>]) -> Result<Vec<Vec<f32>>> {
if inputs.is_empty() {
return Ok(Vec::new());
}
// e5 prefix per §11.3 BEFORE tokenization (same convention as
// FastembedEmbedder so the two backends produce comparable vectors).
let prefixed: Vec<String> = inputs.iter().map(prefix_input).collect();
let mut out: Vec<Vec<f32>> = Vec::with_capacity(prefixed.len());
for chunk in prefixed.chunks(self.batch_size) {
let batch = self.embed_batch(chunk)?;
for v in &batch {
if v.len() != self.dimensions {
anyhow::bail!(
"candle returned vector of length {} but adapter expects {}",
v.len(),
self.dimensions
);
}
}
out.extend(batch);
}
debug_assert_eq!(out.len(), inputs.len());
Ok(out)
}
}
/// Build the e5-prefixed string for one [`EmbeddingInput`]. Free function so
/// a unit test can pin the format without loading the model. Byte-identical to
/// `kebab-embed-local`'s `prefix_input` — the two backends MUST agree here or
/// their vectors diverge.
fn prefix_input(input: &EmbeddingInput<'_>) -> String {
match input.kind {
EmbeddingKind::Document => format!("passage: {}", input.text),
EmbeddingKind::Query => format!("query: {}", input.text),
}
}
/// Apply a one-shot global rayon thread cap (the NUMA-safety lever). Returns
/// `true` if this call set the pool, `false` if it was already initialized
/// (cap not applied) or `n_threads == 0`. `#[doc(hidden)] pub` so the
/// thread-cap test can drive it without loading the 2GB model.
#[doc(hidden)]
pub fn apply_thread_cap(n_threads: usize) -> bool {
if n_threads == 0 {
return false;
}
rayon::ThreadPoolBuilder::new()
.num_threads(n_threads)
.build_global()
.is_ok()
}
/// Compare model hidden size against the configured dim. Extracted so a unit
/// test can exercise the error branch without loading the model.
pub(crate) fn check_dim(model_dim: usize, cfg_dim: usize) -> Result<()> {
if model_dim != cfg_dim {
anyhow::bail!(
"dimension mismatch: model={model_dim}, config={cfg_dim}; \
update `config.models.embedding.dimensions` to match the model \
(or pick a different model)."
);
}
Ok(())
}
#[cfg(test)]
mod tests {
use super::*;
// ── prefix_input ─────────────────────────────────────────────────
// Pin the exact e5 prefix strings; these MUST match
// kebab-embed-local::prefix_input or candle vs fastembed parity breaks.
#[test]
fn prefix_document_uses_passage() {
let input = EmbeddingInput {
text: "hello world",
kind: EmbeddingKind::Document,
};
assert_eq!(prefix_input(&input), "passage: hello world");
}
#[test]
fn prefix_query_uses_query() {
let input = EmbeddingInput {
text: "hello world",
kind: EmbeddingKind::Query,
};
assert_eq!(prefix_input(&input), "query: hello world");
}
#[test]
fn prefix_handles_empty_text() {
let doc = EmbeddingInput {
text: "",
kind: EmbeddingKind::Document,
};
let qry = EmbeddingInput {
text: "",
kind: EmbeddingKind::Query,
};
assert_eq!(prefix_input(&doc), "passage: ");
assert_eq!(prefix_input(&qry), "query: ");
}
// ── check_dim ────────────────────────────────────────────────────
#[test]
fn check_dim_passes_for_1024() {
check_dim(1024, 1024).expect("matching dims must pass");
}
#[test]
fn check_dim_rejects_384_vs_1024() {
let err = check_dim(384, 1024).expect_err("dim mismatch must error");
let msg = format!("{err}");
assert!(
msg.contains("384") && msg.contains("1024"),
"error must mention both dims, got: {msg}"
);
}
}

View File

@@ -0,0 +1,88 @@
//! Parity test (spec §7, `#[ignore]` — needs the ~2GB model + network).
//!
//! Confirms the candle backend reproduces the onnxruntime `FastembedEmbedder`
//! vectors closely enough that no re-index is required (spec D-reindex):
//! per-sentence cosine ≥ 0.9999, and reports the dimension-wise max absolute
//! difference (the number the re-index decision hangs on).
//!
//! Run manually:
//! CARGO_TARGET_DIR=/build/out/cargo-target/target \
//! cargo test -p kebab-embed-candle --release -- --ignored --nocapture
//!
//! Uses the canonical dogfood config so both backends resolve the same model
//! identifiers and cache roots.
use kebab_config::Config;
use kebab_core::{Embedder, EmbeddingInput, EmbeddingKind};
use kebab_embed_candle::CandleEmbedder;
use kebab_embed_local::FastembedEmbedder;
const DOGFOOD_CONFIG: &str = "/build/dogfood/config.toml";
/// Mixed Korean / English parity set (≥ 8 sentences, mirrors the Phase 0 spike).
const SENTENCES: &[&str] = &[
"The quick brown fox jumps over the lazy dog.",
"오늘 날씨가 정말 좋아서 산책을 나가고 싶다.",
"Rust is a systems programming language focused on safety and performance.",
"벡터 검색은 임베딩 사이의 코사인 유사도를 이용한다.",
"Machine learning models require large amounts of training data.",
"한국어와 영어가 섞인 문장도 멀티링구얼 모델은 잘 처리한다.",
"The capital of France is Paris, a city known for its art and culture.",
"이 프로젝트는 로컬 우선 지식 베이스와 검색 증강 생성을 목표로 한다.",
"Database indexing dramatically speeds up query performance.",
"임베딩 모델을 candle 로 옮기면 NUMA 서버에서 안전하게 돌릴 수 있다.",
];
fn cosine(a: &[f32], b: &[f32]) -> f32 {
let dot: f32 = a.iter().zip(b).map(|(x, y)| x * y).sum();
let na: f32 = a.iter().map(|x| x * x).sum::<f32>().sqrt();
let nb: f32 = b.iter().map(|x| x * x).sum::<f32>().sqrt();
dot / (na * nb)
}
#[test]
#[ignore = "needs ~2GB model + network; run manually for the re-index decision"]
fn candle_matches_fastembed() {
let config = Config::load(Some(std::path::Path::new(DOGFOOD_CONFIG)))
.expect("load dogfood config for parity baseline");
let candle = CandleEmbedder::new(&config).expect("build CandleEmbedder");
let fastembed = FastembedEmbedder::new(&config).expect("build FastembedEmbedder");
let inputs: Vec<EmbeddingInput> = SENTENCES
.iter()
.map(|s| EmbeddingInput {
text: s,
kind: EmbeddingKind::Document,
})
.collect();
let cv = candle.embed(&inputs).expect("candle embed");
let fv = fastembed.embed(&inputs).expect("fastembed embed");
assert_eq!(cv.len(), fv.len(), "embedding counts must match");
assert_eq!(candle.dimensions(), 1024);
let mut min_cos = f32::INFINITY;
let mut max_abs_diff = 0f32;
for (i, s) in SENTENCES.iter().enumerate() {
assert_eq!(cv[i].len(), 1024, "candle dim");
assert_eq!(fv[i].len(), 1024, "fastembed dim");
let c = cosine(&cv[i], &fv[i]);
min_cos = min_cos.min(c);
let diff = cv[i]
.iter()
.zip(&fv[i])
.map(|(a, b)| (a - b).abs())
.fold(0f32, f32::max);
max_abs_diff = max_abs_diff.max(diff);
let preview: String = s.chars().take(40).collect();
println!("[{i:>2}] cos={c:.6} max_abs_diff={diff:.6e} {preview}");
}
println!("PARITY_SUMMARY cosine_min={min_cos:.6} max_abs_diff={max_abs_diff:.6e}");
assert!(
min_cos >= 0.9999,
"candle vs fastembed cosine_min={min_cos:.6} < 0.9999 — investigate before merge"
);
}

View File

@@ -0,0 +1,32 @@
//! Thread-cap test (spec §7). Own integration binary → clean process, so the
//! one-shot global rayon pool is initialized exactly once, by us.
//!
//! Verifies that `apply_thread_cap(4)` sizes the global rayon pool to 4, which
//! is the lever that keeps candle's CPU backend NUMA-safe (vs onnxruntime's
//! hard-coded 48 intra-op threads).
use kebab_embed_candle::apply_thread_cap;
#[test]
fn thread_cap_sizes_global_rayon_pool() {
// Must run before any other rayon use in this process. As the only test in
// this binary that touches rayon, that holds.
let applied = apply_thread_cap(4);
assert!(applied, "first build_global call should succeed");
assert_eq!(
rayon::current_num_threads(),
4,
"global rayon pool must be capped at the requested 4 threads"
);
// A second cap attempt is a no-op (pool already built), not a panic.
assert!(
!apply_thread_cap(8),
"second build_global must report not-applied"
);
assert_eq!(
rayon::current_num_threads(),
4,
"thread count must stay at the first cap"
);
}

View File

@@ -1,32 +0,0 @@
# Track 1 / Phase 0 feasibility SPIKE — NOT production.
# Isolated binary that loads multilingual-e5-large via candle (pure Rust)
# and compares its output against the existing onnxruntime FastembedEmbedder.
# candle deps live ONLY here so the production crates stay untouched.
[package]
name = "spike-embed-candle"
version = "0.0.0"
edition = "2024"
publish = false
[[bin]]
name = "spike-embed-candle"
path = "src/main.rs"
[dependencies]
anyhow = "1"
serde_json = "1"
# candle stack — pinned to the current crates.io release (0.10.2).
candle-core = "0.10.2"
candle-nn = "0.10.2"
candle-transformers = "0.10.2"
# Align with workspace-locked versions so we reuse compiled artifacts.
tokenizers = "0.21"
hf-hub = { version = "0.4", features = ["ureq"] }
rayon = "1"
# Parity baseline: reuse the real production embedder + its config loader.
kebab-config = { path = "../kebab-config" }
kebab-embed = { path = "../kebab-embed" }
kebab-embed-local = { path = "../kebab-embed-local" }
# Keep the spike out of the workspace pedantic-lint gate; it is throwaway.
[lints]

View File

@@ -1,251 +0,0 @@
//! Track 1 / Phase 0 feasibility SPIKE (NOT production code).
//!
//! Proves whether candle (pure Rust) can run `intfloat/multilingual-e5-large`
//! with output parity against the existing onnxruntime `FastembedEmbedder`,
//! so the NUMA double-free in fastembed 4.9.1 can be sidestepped.
//!
//! What it checks (see SPIKE_BRIEF.md):
//! 1. numeric parity — per-sentence cosine vs FastembedEmbedder
//! 2. padding_idx — XLM-R position ids start at pad_token_id+1
//! 3. thread control — RAYON_NUM_THREADS caps candle's CPU threads
//! 4. CPU latency — batch wall-clock, rough vs onnxruntime
//!
//! Run:
//! CARGO_TARGET_DIR=/build/out/cargo-target/target \
//! HF_HOME=/build/cache/huggingface \
//! RAYON_NUM_THREADS=4 \
//! cargo run -j 4 -p spike-embed-candle --release
use std::path::PathBuf;
use std::time::Instant;
use anyhow::{Context, Result};
use candle_core::{DType, Device, Tensor};
use candle_nn::VarBuilder;
use candle_transformers::models::xlm_roberta::{Config as XlmConfig, XLMRobertaModel};
use tokenizers::{PaddingParams, PaddingStrategy, Tokenizer, TruncationParams};
use kebab_embed::{Embedder, EmbeddingInput, EmbeddingKind};
use kebab_embed_local::FastembedEmbedder;
const HF_MODEL: &str = "intfloat/multilingual-e5-large";
const DOGFOOD_CONFIG: &str = "/build/dogfood/config.toml";
const MAX_LEN: usize = 512;
/// Mixed Korean / English parity set (≥ 8, brief §3).
const SENTENCES: &[&str] = &[
"The quick brown fox jumps over the lazy dog.",
"오늘 날씨가 정말 좋아서 산책을 나가고 싶다.",
"Rust is a systems programming language focused on safety and performance.",
"벡터 검색은 임베딩 사이의 코사인 유사도를 이용한다.",
"Machine learning models require large amounts of training data.",
"한국어와 영어가 섞인 문장도 멀티링구얼 모델은 잘 처리한다.",
"The capital of France is Paris, a city known for its art and culture.",
"이 프로젝트는 로컬 우선 지식 베이스와 검색 증강 생성을 목표로 한다.",
"Database indexing dramatically speeds up query performance.",
"임베딩 모델을 candle 로 옮기면 NUMA 서버에서 안전하게 돌릴 수 있다.",
];
fn main() -> Result<()> {
// Touch the rayon global pool early so RAYON_NUM_THREADS is honored and
// reportable before any candle compute spins it up.
let rayon_threads = rayon::current_num_threads();
let avail = std::thread::available_parallelism()
.map(|n| n.get())
.unwrap_or(0);
let rayon_env = std::env::var("RAYON_NUM_THREADS").unwrap_or_else(|_| "<unset>".into());
println!("== spike-embed-candle ==");
println!("available_parallelism = {avail}");
println!("RAYON_NUM_THREADS env = {rayon_env}");
println!("rayon::current_num_threads() = {rayon_threads}");
let device = Device::Cpu;
// ── 1. Fetch model files (candle reads safetensors, not the ONNX cache) ──
let cache_dir = std::env::var("HF_HOME")
.map(PathBuf::from)
.unwrap_or_else(|_| PathBuf::from("/build/cache/huggingface"));
let api = hf_hub::api::sync::ApiBuilder::new()
.with_cache_dir(cache_dir.clone())
.build()
.context("build hf-hub api")?;
let repo = api.model(HF_MODEL.to_string());
println!("\n[load] fetching {HF_MODEL} into {} ...", cache_dir.display());
let config_path = repo.get("config.json").context("download config.json")?;
let tokenizer_path = repo.get("tokenizer.json").context("download tokenizer.json")?;
let weights_path = repo
.get("model.safetensors")
.context("download model.safetensors")?;
println!("[load] config = {}", config_path.display());
println!("[load] tokenizer = {}", tokenizer_path.display());
println!("[load] weights = {}", weights_path.display());
// ── 2. Build the candle XLM-RoBERTa model ──
let cfg_json = std::fs::read_to_string(&config_path)?;
let cfg: XlmConfig = serde_json::from_str(&cfg_json).context("parse XLM-R config")?;
println!(
"[load] config: hidden={} layers={} heads={} pad_token_id={} max_pos={} pos_emb={}",
cfg.hidden_size,
cfg.num_hidden_layers,
cfg.num_attention_heads,
cfg.pad_token_id,
cfg.max_position_embeddings,
cfg.position_embedding_type,
);
let vb = unsafe {
VarBuilder::from_mmaped_safetensors(&[weights_path], DType::F32, &device)
.context("mmap safetensors")?
};
let model = XLMRobertaModel::new(&cfg, vb).context("build XLMRobertaModel")?;
let mut tokenizer = Tokenizer::from_file(&tokenizer_path)
.map_err(|e| anyhow::anyhow!("load tokenizer: {e}"))?;
tokenizer
.with_padding(Some(PaddingParams {
strategy: PaddingStrategy::BatchLongest,
..Default::default()
}))
.with_truncation(Some(TruncationParams {
max_length: MAX_LEN,
..Default::default()
}))
.map_err(|e| anyhow::anyhow!("set truncation: {e}"))?;
let pad_id = cfg.pad_token_id;
// ── 3. candle embedding path (passage prefix, masked mean pool, L2) ──
let candle_vecs = candle_embed(&model, &tokenizer, &device, pad_id, SENTENCES)?;
println!("\n[candle] embedded {} sentences, dim={}", candle_vecs.len(), candle_vecs[0].len());
// L2 norm sanity (should be ~1.0 after normalization)
let norm0 = l2(&candle_vecs[0]);
println!("[candle] ‖v0‖ = {norm0:.6}");
// ── 4. FastembedEmbedder (onnxruntime) baseline ──
println!("\n[fastembed] loading FastembedEmbedder from {DOGFOOD_CONFIG} ...");
let config = kebab_config::Config::load(Some(std::path::Path::new(DOGFOOD_CONFIG)))
.context("load dogfood config")?;
let fb_t0 = Instant::now();
let fb = FastembedEmbedder::new(&config).context("build FastembedEmbedder")?;
println!("[fastembed] model loaded in {:.2}s", fb_t0.elapsed().as_secs_f64());
let fb_inputs: Vec<EmbeddingInput> = SENTENCES
.iter()
.map(|s| EmbeddingInput { text: s, kind: EmbeddingKind::Document })
.collect();
let fb_vecs = fb.embed(&fb_inputs).context("fastembed embed")?;
// ── 5. Per-sentence parity (both L2-normalized → cosine = dot) ──
println!("\n== PARITY (candle vs fastembed, EmbeddingKind::Document / passage:) ==");
let mut cosines = Vec::with_capacity(SENTENCES.len());
for (i, s) in SENTENCES.iter().enumerate() {
let c = cosine(&candle_vecs[i], &fb_vecs[i]);
cosines.push(c);
let preview: String = s.chars().take(40).collect();
println!(" [{i:>2}] cos={c:.6} {preview}");
}
let min = cosines.iter().cloned().fold(f32::INFINITY, f32::min);
let mean = cosines.iter().sum::<f32>() / cosines.len() as f32;
println!(" --> cosine min={min:.6} mean={mean:.6}");
// ── 6. Latency: batch of 32 (replicated) through candle ──
let batch: Vec<&str> = SENTENCES.iter().cloned().cycle().take(32).collect();
// warmup
let _ = candle_embed(&model, &tokenizer, &device, pad_id, &batch[..4])?;
let t0 = Instant::now();
let _ = candle_embed(&model, &tokenizer, &device, pad_id, &batch)?;
let candle_lat = t0.elapsed();
let fb_batch: Vec<EmbeddingInput> = batch
.iter()
.map(|s| EmbeddingInput { text: s, kind: EmbeddingKind::Document })
.collect();
let t1 = Instant::now();
let _ = fb.embed(&fb_batch)?;
let fb_lat = t1.elapsed();
let peak_threads = proc_threads();
println!("\n== LATENCY (batch=32) ==");
println!(" candle : {:.3}s ({:.1} ms/sentence)", candle_lat.as_secs_f64(), candle_lat.as_secs_f64() * 1000.0 / 32.0);
println!(" fastembed : {:.3}s ({:.1} ms/sentence)", fb_lat.as_secs_f64(), fb_lat.as_secs_f64() * 1000.0 / 32.0);
println!("\n== THREAD CONTROL ==");
println!(" RAYON_NUM_THREADS env = {rayon_env}");
println!(" rayon::current_num_threads = {rayon_threads}");
println!(" available_parallelism = {avail}");
println!(" peak OS threads (/proc) = {peak_threads}");
// ── 7. Machine verdict line for the report ──
let verdict = if mean >= 0.99 { "PASS" } else if mean >= 0.95 { "MARGINAL" } else { "FAIL" };
println!("\n== SUMMARY ==");
println!("VERDICT_HINT={verdict} cosine_min={min:.6} cosine_mean={mean:.6} candle_batch32_s={:.3} fb_batch32_s={:.3} rayon_threads={rayon_threads} rayon_env={rayon_env}", candle_lat.as_secs_f64(), fb_lat.as_secs_f64());
Ok(())
}
/// candle embedding: apply e5 `passage:` prefix, tokenize (batch-padded),
/// forward through XLM-R, attention-mask-weighted mean pool, L2 normalize.
fn candle_embed(
model: &XLMRobertaModel,
tokenizer: &Tokenizer,
device: &Device,
_pad_id: u32,
sentences: &[&str],
) -> Result<Vec<Vec<f32>>> {
let prefixed: Vec<String> = sentences.iter().map(|s| format!("passage: {s}")).collect();
let encodings = tokenizer
.encode_batch(prefixed, true)
.map_err(|e| anyhow::anyhow!("encode_batch: {e}"))?;
let bsz = encodings.len();
let seq = encodings[0].get_ids().len();
let mut ids = Vec::with_capacity(bsz * seq);
let mut mask = Vec::with_capacity(bsz * seq);
for enc in &encodings {
ids.extend(enc.get_ids().iter().copied());
mask.extend(enc.get_attention_mask().iter().map(|&m| m as f32));
}
let input_ids = Tensor::from_vec(ids, (bsz, seq), device)?;
let attn_f32 = Tensor::from_vec(mask, (bsz, seq), device)?;
let token_type_ids = input_ids.zeros_like()?;
// forward: (input_ids, attention_mask, token_type_ids, past, enc_hidden, enc_mask)
let hidden = model.forward(&input_ids, &attn_f32, &token_type_ids, None, None, None)?;
// masked mean pool
let mask3 = attn_f32.unsqueeze(2)?; // (b, seq, 1)
let summed = hidden.broadcast_mul(&mask3)?.sum(1)?; // (b, hidden)
let counts = mask3.sum(1)?; // (b, 1)
let mean = summed.broadcast_div(&counts)?;
// L2 normalize
let norm = mean.sqr()?.sum_keepdim(1)?.sqrt()?;
let normalized = mean.broadcast_div(&norm)?;
Ok(normalized.to_vec2::<f32>()?)
}
fn cosine(a: &[f32], b: &[f32]) -> f32 {
let dot: f32 = a.iter().zip(b).map(|(x, y)| x * y).sum();
let na = l2(a);
let nb = l2(b);
dot / (na * nb)
}
fn l2(v: &[f32]) -> f32 {
v.iter().map(|x| x * x).sum::<f32>().sqrt()
}
/// Peak OS thread count for this process from /proc/self/status.
fn proc_threads() -> usize {
std::fs::read_to_string("/proc/self/status")
.ok()
.and_then(|s| {
s.lines()
.find(|l| l.starts_with("Threads:"))
.and_then(|l| l.split_whitespace().nth(1).map(str::to_string))
})
.and_then(|n| n.parse().ok())
.unwrap_or(0)
}

View File

@@ -66,7 +66,8 @@ flowchart TB
end
subgraph Adapters ["traits + adapters"]
embed["kebab-embed<br/>(trait)"]
embedlocal["kebab-embed-local<br/>(fastembed)"]
embedlocal["kebab-embed-local<br/>(fastembed, default)"]
embedcandle["kebab-embed-candle<br/>(candle, NUMA-safe opt-in)"]
llm["kebab-llm<br/>(trait)"]
llmlocal["kebab-llm-local<br/>(Ollama)"]
search["kebab-search"]
@@ -92,6 +93,7 @@ flowchart TB
app --> sqlite
app --> vector
app --> embedlocal
app --> embedcandle
app --> llmlocal
app --> search
app --> rag
@@ -104,6 +106,8 @@ flowchart TB
paud --> core
pcode --> core
embedlocal --> embed
embedcandle --> core
embedcandle --> config
llmlocal --> llm
rag --> search
rag --> llm
@@ -180,6 +184,7 @@ kebab/
│ ├── kebab-store-sqlite/ # SQLite + FTS5 (V001/V002/V003) (P1-6, P2-1, P3-3). src/derivation_cache.rs = derivation_cache 테이블 저장소 (V012, v0.21.0)
│ ├── kebab-search/ # Lexical + Vector + Hybrid retriever (P2-2, P3-4)
│ ├── kebab-embed/ kebab-embed-local/ # Embedder trait + fastembed adapter (P3-1, P3-2)
│ ├── kebab-embed-candle/ # candle (pure-Rust) Embedder, NUMA-safe opt-in provider=candle (Track 1, v0.22.0)
│ ├── kebab-store-vector/ # LanceDB VectorStore (P3-3, P7-3 follow-up)
│ ├── kebab-llm/ kebab-llm-local/ # LanguageModel trait + Ollama adapter (P4-1, P4-2)
│ ├── kebab-rag/ # RAG pipeline (P4-3)

View File

@@ -107,11 +107,13 @@ respect_markdown_headings = true
chunker_version = "md-heading-v1"
[models.embedding]
provider = "fastembed" # "none" 으로 두면 lexical-only — Ollama 불필요
provider = "fastembed" # "fastembed"(기본) / "candle"(순수 Rust, NUMA-안전)
# / "none"(lexical-only — Ollama 불필요)
model = "multilingual-e5-small"
version = "v1"
dimensions = 384
batch_size = 64
num_threads = 0 # candle 전용 CPU 스레드 캡 (0=auto). env KEBAB_EMBED_THREADS 우선.
[models.llm]
provider = "ollama"

View File

@@ -0,0 +1,72 @@
---
title: kebab v0.22.0 release notes (draft)
created: 2026-06-01
status: draft
release_trigger:
- 신규 config surface (provider=candle, num_threads / KEBAB_EMBED_THREADS) — pre-1.0 minor bump
- 임베딩 백엔드 다변화 (NUMA-안전 candle provider 추가, opt-in)
---
# kebab v0.22.0 — candle 임베딩 provider (NUMA-안전, opt-in)
v0.21.1 (config 마이그레이션) 후속 minor release. 듀얼소켓 NUMA 서버에서
onnxruntime 의 스레드 하드코딩이 일으키던 ingest 크래시를 피하기 위해, 같은
임베딩 모델을 **순수 Rust(candle)** 로 돌리는 opt-in provider 를 추가한다.
**기본 동작은 그대로다** — 기존 사용자는 아무것도 바꿀 필요가 없다.
---
## 핵심 변경
### candle 임베딩 provider (`provider = "candle"`)
**변경 사실.** `[models.embedding].provider``"candle"` 값이 추가됐다.
`"fastembed"`(기본, onnxruntime) / `"candle"`(순수 Rust) / `"none"`(lexical-only)
중 하나를 고를 수 있다. candle provider 는 fastembed 와 **완전히 같은 모델**
(`intfloat/multilingual-e5-large`, 1024-dim)을 쓰고, e5 prefix → mean pooling
→ L2 정규화 파이프라인도 동일하다. 첫 사용 시 safetensors(~2GB)를
`{model_dir}/candle/` 아래로 자동 다운로드한다.
```toml
[models.embedding]
provider = "candle" # 기본은 "fastembed" — NUMA 서버에서만 candle 권장
num_threads = 8 # candle CPU 스레드 캡 (0 = auto = #cores)
```
```bash
# env 로도 캡 가능 (config 보다 우선)
KEBAB_EMBED_THREADS=8 kebab ingest
```
**Trade-off.** candle 는 순수 Rust 라 onnxruntime 의 네이티브 SIMD 커널보다
CPU latency 가 느리다 (Phase 0 스파이크 측정 ~4×). 그래서 **기본값은
fastembed 를 유지**하고, candle 은 onnxruntime 가 죽는 NUMA 환경에서만 켜는
opt-in 으로 둔다. 단일 워크스테이션 사용자는 fastembed 가 더 빠르다.
**Mitigation (왜 안전한가).** candle 의 CPU 백엔드는 글로벌 rayon 풀 크기로
스레드를 정한다. `num_threads`(또는 env `KEBAB_EMBED_THREADS`)가 그 풀을 한 번
캡하므로, onnxruntime 가 하드코딩하던 48 intra-op 스레드 → NUMA 힙 손상 →
double-free 경로를 원천 차단한다. NUMA 노드 바인딩이 더 필요하면 `numactl`
과 조합한다.
**Upgrade 절차.** 재색인 **불필요**. candle 과 fastembed 의 벡터는 사실상
동일(Phase 0 스파이크 코사인 1.000000)해서 `embedding_version` 을 유지했고,
기존 LanceDB 색인을 그대로 재사용한다. provider 를 바꿔도 검색 결과는
바뀌지 않는다. 기존 `config.toml``num_threads` 가 자동으로 `0`(auto)으로
채워져 그대로 로드된다 — `kebab config migrate` 도, 수동 편집도 필요 없다.
---
## 그 외
- 신규 crate `kebab-embed-candle` (candle 의존성 트리를 이 crate 에 격리,
`kebab-core`/`kebab-config` 외 다른 kebab-* 의존 없음).
- Phase 0 feasibility 스파이크(`spike-embed-candle`)는 production 흡수 후 제거.
- 문서: README Configuration, `docs/SMOKE.md` config 예시, `docs/ARCHITECTURE.md`
crate 그래프/트리에 candle provider 반영.
## 잔여 검증 (사용자 실행)
듀얼소켓 NUMA 서버에서 `provider=candle` 로 5150-doc ingest 가 double-free
없이 EXIT=0 완주하는지가 본 release 의 최종 인수 게이트다 (meta-spec §4.3).
패리티 max abs diff 수치는 `IMPL_REPORT.md` 참조.

View File

@@ -14,6 +14,52 @@ historical contract that was implemented; this file accumulates the
deltas so phase 5+ readers can find the live behavior without diffing
git history.
## 2026-06-01 — candle 임베딩 provider (NUMA double-free 회피, opt-in)
**무엇이 문제였나.** 듀얼소켓 NUMA 서버에서 `provider=fastembed`(onnxruntime)로
대규모 ingest(5150-doc)를 돌리면 onnxruntime 가 intra-op 스레드를 48개로
하드코딩해 NUMA 힙을 손상시키고 double-free 로 프로세스가 죽었다. 스레드 수를
config 로 줄일 surface 가 없었고, fastembed 4.9 의 ORT 바인딩은 이를 노출하지
않는다.
**진단 / 결정 (사용자 승인 2026-06-01).** 같은 모델
`intfloat/multilingual-e5-large`**candle(순수 Rust)** 로 돌리는 임베딩
provider 를 추가하기로 결정. candle 의 CPU 백엔드는 글로벌 rayon 풀 크기로
스레드를 정하므로, 한 번의 `rayon::ThreadPoolBuilder::build_global` 캡으로
스레드를 NUMA-안전한 수로 묶을 수 있다. **재색인 0 목표**(`embedding_version`
유지) — Phase 0 스파이크(`SPIKE_REPORT.md`, 커밋 76841af)가 candle vs
onnxruntime **코사인 1.000000** 패리티를 입증했고, 본 Track 1 구현의 패리티
테스트로 차원별 max 절대오차를 재실측해 확정.
**무엇을 건드렸나.**
- 신규 crate `crates/kebab-embed-candle``kebab_core::Embedder` 구현
(`CandleEmbedder`). 스파이크 파이프라인(safetensors via hf-hub → XLM-RoBERTa
forward → attention-mask mean pooling → L2 → e5 prefix)을 production 으로
흡수. deps 는 candle 트리를 이 crate 에 격리 (core/config 외 다른 kebab-*
의존 0 — design §8 경계). 모델 캐시 `{model_dir}/candle/`.
- 스레드 캡: `[models.embedding].num_threads`(u32, default 0=auto) + env
`KEBAB_EMBED_THREADS`(우선). `CandleEmbedder::new` 에서 n>0 이면 글로벌 rayon
풀 1회 캡(이미 init 시 no-op).
- 주입 분기: `kebab-app::App::embedder()``config.models.embedding.provider`
분기 — `fastembed`/`onnx`/(빈값) → 기존 `FastembedEmbedder`(동작 불변),
`candle``CandleEmbedder`, 미지값 → 에러. `none` 은 기존 lexical-only 유지.
- 스파이크 crate `crates/spike-embed-candle` 제거(학습은 production 으로 흡수됨).
- 버전 0.21.1 → **0.22.0** (신규 config surface — pre-1.0 minor bump).
**패리티 증거.** Phase 0 스파이크 cosine 1.000000 (10문장 한/영 혼합). 본
Track 1 의 `#[ignore]` 패리티 테스트 결과(max abs diff)는
`/build/out/kebab-worktrees/embed-candle/IMPL_REPORT.md` 에 기록.
**호환성.** fastembed default 경로의 동작/벡터 불변. `embedding_version`
유지 → 기존 색인 재사용(재색인 0). wire schema 변경 없음. 옛 config.toml 은
`num_threads` 가 serde default(0)로 채워져 그대로 파싱.
**잔여 게이트 (사용자 실행, Claude 불가).** 그 듀얼소켓 NUMA 서버에서
`provider=candle` 로 5150-doc ingest 가 double-free 없이 EXIT=0 완주하는지
PR 머지 전/후 검증 예약 (meta-spec §4.3).
amends: `docs/superpowers/specs/2026-06-01-embed-candle-track-spec.md`.
## 2026-05-31 — config 마이그레이션 (`kebab config migrate`)
**Trigger**: config.toml 스키마가 진화해도(v0.21.0 의 `[ingest.expansion]` 등) 기존 사용자 파일은 serde default 로 *동작*만 호환될 뿐 새 섹션이 파일에 안 써져 사용자가 노브의 존재를 알 수 없었다. DB 의 V00X refinery 와 달리 config 엔 마이그레이션 메커니즘이 없어 추가. 설계 `docs/superpowers/specs/2026-05-31-config-migration-design.md`, 계획 `docs/superpowers/plans/2026-05-31-config-migration.md`, PR #198.