# WARM comparison
Generated by `./benchmarks/competitive/warm-bench table` from the committed files in `benchmarks/results/competitive/`; do not edit by hand. Every cell is indexed with its source file and key path in [comparison.cells.json](comparison.cells.json). Method, caveats and how to reproduce: [README.md](README.md).
Rows are never compared across hardware classes: each section is one dataset on one hardware class. A verdict is WIN or LOSS only when the 95% bootstrap interval of the advantage factor excludes 1 (deterministic byte counts use a 5% band); otherwise TIE. The advantage factor is always oriented so that more than 1 means WARM is better. Hosted-runner rows carry noise; trust the ratios, not the absolute values.
## Scoreboard
| Dataset | Hardware class | Suite | WARM wins | Ties | Losses | Not measured | Informational (no verdict) |
|---|---|---|---|---|---|---|---|
| sift-128-euclidean (1,000,000 vectors, 10,000 queries) | `linux-x86_64-intel-ultra-9-185h-22c-27g` | heavy | 0 | 0 | 0 | 0 | 14 |
| beir:fiqa (57,600 documents, 500 queries) | `linux-x86_64-intel-ultra-9-185h-22c-27g` | standard | 0 | 0 | 0 | 2 | 12 |
| memory (4,559 documents, 100 queries) | `modal-x86_64-8vcpu-32g` | heavy | 1 | 5 | 4 | 0 | 4 |
| memory (4,559 documents, 100 queries) | `linux-x86_64-intel-ultra-9-185h-22c-27g` | standard | 0 | 0 | 0 | 2 | 12 |
| beir:nfcorpus (3,633 documents, 324 queries) | `linux-x86_64-intel-ultra-9-185h-22c-27g` | standard | 0 | 0 | 0 | 2 | 12 |
| beir:nfcorpus (3,633 documents, 324 queries) | `modal-x86_64-8vcpu-32g` | heavy | 2 | 3 | 5 | 0 | 4 |
| beir:scifact (5,183 documents, 809 queries) | `linux-x86_64-intel-ultra-9-185h-22c-27g` | standard | 0 | 0 | 0 | 0 | 14 |
| beir:scifact (5,183 documents, 809 queries) | `modal-x86_64-8vcpu-32g` | heavy | 0 | 3 | 7 | 0 | 4 |
## Where WARM loses
| Dataset | Metric | Best other | WARM disadvantage | Candidate lever (hypothesis unless a file is cited) |
|---|---|---|---|---|
| memory (4,559 documents, 100 queries) (modal-x86_64-8vcpu-32g) | Ingest/build time | SQLite FTS5 | 1.22x worse | unprofiled hypothesis: per-document text encoding in documents::prepare (src/core/encoding.rs) dominates ingest; profile it, then a lexical-only ingest path or a cheaper encoder is the lever |
| memory (4,559 documents, 100 queries) (modal-x86_64-8vcpu-32g) | Peak RSS, build | SQLite FTS5 | 1.10x worse | unprofiled hypothesis: the arena and index are built in memory before commit; streaming commits per batch |
| memory (4,559 documents, 100 queries) (modal-x86_64-8vcpu-32g) | Peak RSS, serving | SQLite FTS5 | 2.55x worse | the whole store is resident after open; memory-mapped postings and file-backed text (ledger 2A/2B) are the open levers |
| memory (4,559 documents, 100 queries) (modal-x86_64-8vcpu-32g) | Cold start to first answer | tantivy | 8.63x worse | the store is read into memory at open (open_s in the raw file); lazy or memory-mapped loading of the arena and postings is the lever |
| beir:nfcorpus (3,633 documents, 324 queries) (modal-x86_64-8vcpu-32g) | Ingest/build time | SQLite FTS5 | 2.07x worse | unprofiled hypothesis: per-document text encoding in documents::prepare (src/core/encoding.rs) dominates ingest; profile it, then a lexical-only ingest path or a cheaper encoder is the lever |
| beir:nfcorpus (3,633 documents, 324 queries) (modal-x86_64-8vcpu-32g) | Peak RSS, build | SQLite FTS5 | 1.69x worse | unprofiled hypothesis: the arena and index are built in memory before commit; streaming commits per batch |
| beir:nfcorpus (3,633 documents, 324 queries) (modal-x86_64-8vcpu-32g) | Peak RSS, serving | SQLite FTS5 | 1.90x worse | the whole store is resident after open; memory-mapped postings and file-backed text (ledger 2A/2B) are the open levers |
| beir:nfcorpus (3,633 documents, 324 queries) (modal-x86_64-8vcpu-32g) | On-disk size | Apache Lucene 9.12.3 (JVM) | 1.39x worse | compact postings and file-backed text (ledger 2B/2C) and removing preallocated arena slack |
| beir:nfcorpus (3,633 documents, 324 queries) (modal-x86_64-8vcpu-32g) | Cold start to first answer | tantivy | 2.57x worse | the store is read into memory at open (open_s in the raw file); lazy or memory-mapped loading of the arena and postings is the lever |
| beir:scifact (5,183 documents, 809 queries) (modal-x86_64-8vcpu-32g) | QPS (single thread) | tantivy | 1.08x worse | |
| beir:scifact (5,183 documents, 809 queries) (modal-x86_64-8vcpu-32g) | p50 latency | tantivy | 1.12x worse | none identified; the median is within noise of the leader |
| beir:scifact (5,183 documents, 809 queries) (modal-x86_64-8vcpu-32g) | Ingest/build time | SQLite FTS5 | 3.32x worse | unprofiled hypothesis: per-document text encoding in documents::prepare (src/core/encoding.rs) dominates ingest; profile it, then a lexical-only ingest path or a cheaper encoder is the lever |
| beir:scifact (5,183 documents, 809 queries) (modal-x86_64-8vcpu-32g) | Peak RSS, build | SQLite FTS5 | 1.67x worse | unprofiled hypothesis: the arena and index are built in memory before commit; streaming commits per batch |
| beir:scifact (5,183 documents, 809 queries) (modal-x86_64-8vcpu-32g) | Peak RSS, serving | SQLite FTS5 | 2.93x worse | the whole store is resident after open; memory-mapped postings and file-backed text (ledger 2A/2B) are the open levers |
| beir:scifact (5,183 documents, 809 queries) (modal-x86_64-8vcpu-32g) | On-disk size | Apache Lucene 9.12.3 (JVM) | 1.47x worse | compact postings and file-backed text (ledger 2B/2C) and removing preallocated arena slack |
| beir:scifact (5,183 documents, 809 queries) (modal-x86_64-8vcpu-32g) | Cold start to first answer | SQLite FTS5 | 3.12x worse | the store is read into memory at open (open_s in the raw file); lazy or memory-mapped loading of the arena and postings is the lever |
## Results
### sift-128-euclidean (1,000,000 vectors, 10,000 queries)
Suite `heavy`, hardware class `linux-x86_64-intel-ultra-9-185h-22c-27g` (Intel(R) Core(TM) Ultra 9 185H, 22 logical cores, 27 GiB, Linux 6.18.33.2-microsoft-standard-WSL2), commit `unknown`, 2026-10-21. Raw results: [heavy-ann-sift-128-euclidean-n1000000-linux-x86_64-intel-ultra-9-185h-22c-27g.json](../../benchmarks/results/competitive/heavy-ann-sift-128-euclidean-n1000000-linux-x86_64-intel-ultra-9-185h-22c-27g.json). Intervals: 95.00% bootstrap (Bonferroni family of 0 verdict metrics).
Timing boundary (where each query is timed): hnswlib: CPython binding call per query; WARM: native Rust API call in the engine process (no language binding); text: BM25 index only (`--encode lexical`); ANN: public-bench loop.
| Metric | WARM | Best other | Advantage (WARM vs best other) | Verdict |
|---|---|---|---|---|
| QPS @ recall@10 >= 0.90 (QPS) | 23,294† | hnswlib 9,480† | n/v (no CI) | **n/v** |
| QPS @ recall@10 >= 0.95 (QPS) | 17,481† | hnswlib 7,630† | n/v (no CI) | **n/v** |
| QPS @ recall@10 >= 0.99 (QPS) | 8,794† | hnswlib 3,193† | n/v (no CI) | **n/v** |
| p50 latency @ 0.95 (us) | 62† | hnswlib 133† | n/v (no CI) | **n/v** |
| p99 latency @ 0.95 (us) | 105† | hnswlib 220† | n/v (no CI) | **n/v** |
| Build time (s) | 312.60† | hnswlib 182.49† | n/v (no CI) | **n/v** |
| Peak RSS, build (MiB) | 2,968† | hnswlib 1,525† | n/v (no CI) | **n/v** |
| RSS, serving (MiB) | 2,204† | hnswlib 1,019† | n/v (no CI) | **n/v** |
| Index size (MiB) | 446.3† | hnswlib 873.8† | n/v (no CI) | **n/v** |
| Cold start to first answer (s) | n/m† (not measured) | hnswlib 0.75† | n/m | **n/v** |
| Replace one vector (durable) (ms) | n/m† (VGraph is a static graph index built once; the document store handles updates with an exact scan, not an ANN index) | hnswlib 6.010† | n/m | **n/v** |
| Delete one vector (durable) (ms) | n/m† (VGraph is a static graph index built once; the document store handles updates with an exact scan, not an ANN index) | hnswlib 3.314† | n/m | **n/v** |
| Install size (MiB) | 3.4† | hnswlib 61.8† | n/v (no CI) | **n/v** |
| Install distributions (count) | 92† | hnswlib 2† | n/v (no CI) | **n/v** |
<details><summary>All systems</summary>
| Metric | hnswlib | WARM |
|---|---|---|
| QPS @ recall@10 >= 0.90 (QPS) | 9,480† | 23,294 |
| QPS @ recall@10 >= 0.95 (QPS) | 7,630† | 17,481 |
| QPS @ recall@10 >= 0.99 (QPS) | 3,193† | 8,794 |
| p50 latency @ 0.95 (us) | 133† | 62 |
| p99 latency @ 0.95 (us) | 220† | 105 |
| Build time (s) | 182.49† | 312.60 |
| Peak RSS, build (MiB) | 1,525† | 2,968 |
| RSS, serving (MiB) | 1,019† | 2,204 |
| Index size (MiB) | 873.8† | 446.3 |
| Cold start to first answer (s) | 0.75† | n/m (not measured) |
| Replace one vector (durable) (ms) | 6.010† | n/m (VGraph is a static graph index built once; the document store handles updates with an exact scan, not an ANN index) |
| Delete one vector (durable) (ms) | 3.314† | n/m (VGraph is a static graph index built once; the document store handles updates with an exact scan, not an ANN index) |
| Install size (MiB) | 61.8† | 3.4 |
| Install distributions (count) | 2† | 92 |
Bold marks the best value in the row. Operating points for the QPS rows: hnswlib: m48e500 {'ef': 32} recall 0.961; WARM: vertex-d32 {'ef': 24, 'patience': 0, 'rerank': 32, 'screen': False, 'screen_sigmas': 1.0} recall 0.957.
</details>
† WARM vs hnswlib: source revision is not pinned; no verdict.
### beir:fiqa (57,600 documents, 500 queries)
Suite `standard`, hardware class `linux-x86_64-intel-ultra-9-185h-22c-27g` (Intel(R) Core(TM) Ultra 9 185H, 22 logical cores, 27 GiB, Linux 6.18.33.2-microsoft-standard-WSL2), commit `unknown`, 2026-10-21. Raw results: [standard-text-beir-fiqa-linux-x86_64-intel-ultra-9-185h-22c-27g.json](../../benchmarks/results/competitive/standard-text-beir-fiqa-linux-x86_64-intel-ultra-9-185h-22c-27g.json). Intervals: 95.00% bootstrap (Bonferroni family of 0 verdict metrics).
Timing boundary (where each query is timed): WARM: native Rust API call in the engine process (no language binding); text: BM25 index only (`--encode lexical`); ANN: public-bench loop; SQLite FTS5: CPython sqlite3 call per query (query string built outside the timer); Apache Lucene 9.12.3 (JVM): in-process JVM API call, after a full untimed JIT warm-up pass; tantivy: CPython tantivy-py call per query (stored id fetched per hit).
| Metric | WARM | Best other | Advantage (WARM vs best other) | Verdict |
|---|---|---|---|---|
| nDCG@10 | 0.2654† | SQLite FTS5 0.2678† | n/v (no CI) | **n/v** |
| recall@100 | 0.5422† | SQLite FTS5 0.5546† | n/v (no CI) | **n/v** |
| QPS (single thread) (QPS) | 3,612† | tantivy 1,617† | n/v (no CI) | **n/v** |
| p50 latency (us) | 269† | tantivy 600† | n/v (no CI) | **n/v** |
| p99 latency (us) | 509† | tantivy 1,269† | n/v (no CI) | **n/v** |
| Ingest/build time (s) | 2.16† | SQLite FTS5 1.18† | n/v (no CI) | **n/v** |
| Peak RSS, build (MiB) | n/m (Linux child inherited the orchestrator high-water mark; measured before the isolated-spawner fix) | n/m | n/m | **n/m** |
| Peak RSS, serving (MiB) | n/m (Linux child inherited the orchestrator high-water mark; measured before the isolated-spawner fix) | n/m | n/m | **n/m** |
| On-disk size (MiB) | 20.6† | Apache Lucene 9.12.3 (JVM) 12.2† | n/v (no CI) | **n/v** |
| Cold start to first answer (s) | 0.84† | tantivy 0.02† | n/v (no CI) | **n/v** |
| Replace one document (durable) (ms) | 1.719† | SQLite FTS5 9.142† | n/v (no CI) | **n/v** |
| Delete one document (durable) (ms) | 3.756† | SQLite FTS5 8.763† | n/v (no CI) | **n/v** |
| Install size (MiB) | 2.5† | SQLite FTS5 0.0† | n/v (no CI) | **n/v** |
| Install distributions (count) | 92† | SQLite FTS5 0† | n/v (no CI) | **n/v** |
<details><summary>All systems</summary>
| Metric | WARM | SQLite FTS5 | Apache Lucene 9.12.3 (JVM) | tantivy |
|---|---|---|---|---|
| nDCG@10 | 0.2654 | 0.2678† | 0.2657† | 0.2643† |
| recall@100 | 0.5422 | 0.5546† | 0.5445† | 0.5406† |
| QPS (single thread) (QPS) | 3,612 | 26† | 1,268† | 1,617† |
| p50 latency (us) | 269 | 34,692† | 721† | 600† |
| p99 latency (us) | 509 | 88,406† | 1,755† | 1,269† |
| Ingest/build time (s) | 2.16 | 1.18† | 2.63† | 2.65† |
| Peak RSS, build (MiB) | n/m (Linux child inherited the orchestrator high-water mark; measured before the isolated-spawner fix) | n/m (Linux child inherited the orchestrator high-water mark; measured before the isolated-spawner fix) | n/m (Linux child inherited the orchestrator high-water mark; measured before the isolated-spawner fix) | n/m (Linux child inherited the orchestrator high-water mark; measured before the isolated-spawner fix) |
| Peak RSS, serving (MiB) | n/m (Linux child inherited the orchestrator high-water mark; measured before the isolated-spawner fix) | n/m (Linux child inherited the orchestrator high-water mark; measured before the isolated-spawner fix) | n/m (Linux child inherited the orchestrator high-water mark; measured before the isolated-spawner fix) | n/m (Linux child inherited the orchestrator high-water mark; measured before the isolated-spawner fix) |
| On-disk size (MiB) | 20.6 | 27.1† | 12.2† | 17.4† |
| Cold start to first answer (s) | 0.84 | 0.08† | 0.29† | 0.02† |
| Replace one document (durable) (ms) | 1.719 | 9.142† | 31.038† | 82.711† |
| Delete one document (durable) (ms) | 3.756 | 8.763† | 14.948† | 43.996† |
| Install size (MiB) | 2.5 | 0.0† | 5.7† | 13.8† |
| Install distributions (count) | 92 | 0† | 2† | 1† |
Bold marks the best value in the row.
</details>
† WARM vs SQLite FTS5: source revision is not pinned; no verdict.
† WARM vs tantivy: source revision is not pinned; no verdict.
† WARM vs Apache Lucene 9.12.3 (JVM): source revision is not pinned; no verdict.
### memory (4,559 documents, 100 queries)
Suite `heavy`, hardware class `modal-x86_64-8vcpu-32g` (unknown, 24 logical cores, 396 GiB, Linux 4.19.0-gvisor), commit `d5f3f3b8cb`, 2026-10-21. Raw results: [heavy-text-memory-modal-x86_64-8vcpu-32g.json](../../benchmarks/results/competitive/heavy-text-memory-modal-x86_64-8vcpu-32g.json). Intervals: 99.50% bootstrap (Bonferroni family of 10 verdict metrics).
Timing boundary (where each query is timed): WARM: native Rust API call in the engine process (no language binding); text: BM25 index only (`--encode lexical`); ANN: public-bench loop; WARM (eager hypervectors): native Rust API call in the engine process (no language binding); text: BM25 index plus chunk hypervectors at ingest (`--encode eager`); SQLite FTS5: CPython sqlite3 call per query (query string built outside the timer); Apache Lucene 9.12.3 (JVM): in-process JVM API call, after a full untimed JIT warm-up pass; tantivy: CPython tantivy-py call per query (stored id fetched per hit); Whoosh: pure Python; bm25s: CPython + NumPy (single-threaded retrieve).
| Metric | WARM | Best other | Advantage (WARM vs best other) | Verdict |
|---|---|---|---|---|
| nDCG@10 | 0.4561 | Apache Lucene 9.12.3 (JVM) 0.4627 | 0.99x [0.60, 1.65] | **TIE** |
| recall@100 | 0.8203 | SQLite FTS5 0.8392 | 0.98x [0.79, 1.22] | **TIE** |
| QPS (single thread) (QPS) | 2,289 | tantivy 2,504 | 0.91x [0.83, 1.04] | **TIE** |
| p50 latency (us) | 420 | tantivy 380 | 0.90x [0.83, 1.04] | **TIE** |
| p99 latency (us) | 775 | tantivy 720 | 0.93x [0.86, 1.02] | **TIE** |
| Ingest/build time (s) | 2.52 | SQLite FTS5 2.06 | 0.82x [0.79, 0.85] | **LOSS** |
| Peak RSS, build (MiB) | 133 | SQLite FTS5 121 | 0.91x [0.91, 0.91] | **LOSS** |
| Peak RSS, serving (MiB) | 66 | SQLite FTS5 26 | 0.39x [0.38, 0.41] | **LOSS** |
| On-disk size (MiB) | 9.8 | Apache Lucene 9.12.3 (JVM) 10.3 | 1.05x (no CI) | **WIN** |
| Cold start to first answer (s) | 0.62 | tantivy 0.07 | 0.12x [0.10, 0.13] | **LOSS** |
| Replace one document (durable) (ms) | 3.919 | SQLite FTS5 4.626 | 1.18x (no CI) | **n/v** |
| Delete one document (durable) (ms) | 3.162 | Apache Lucene 9.12.3 (JVM) 1.482 | 0.47x (no CI) | **n/v** |
| Install size (MiB) | 2.4 | Whoosh 2.0 | 0.83x (no CI) | **n/v** |
| Install distributions (count) | 92 | tantivy 1 | 0.01x (no CI) | **n/v** |
<details><summary>All systems</summary>
| Metric | WARM | WARM (eager hypervectors) | SQLite FTS5 | Apache Lucene 9.12.3 (JVM) | tantivy | Whoosh | bm25s |
|---|---|---|---|---|---|---|---|
| nDCG@10 | 0.4561 | 0.4561 | 0.4501 | **0.4627** | 0.4491 | 0.4401 | 0.4418 |
| recall@100 | 0.8203 | 0.8203 | **0.8392** | 0.8237 | 0.8203 | 0.8207 | 0.8312 |
| QPS (single thread) (QPS) | 2,289 | 2,232 | 32 | 1,002 | **2,504** | 14 | 1,102 |
| p50 latency (us) | 420 | 432 | 24,740 | 930 | **380** | 68,913 | 900 |
| p99 latency (us) | 775 | 780 | 118,386 | 2,213 | **720** | 131,749 | 1,249 |
| Ingest/build time (s) | 2.52 | 16.57 | **2.06** | 4.09 | 3.48 | 41.11 | 5.15 |
| Peak RSS, build (MiB) | 133 | 167 | **121** | 768 | 244 | 536 | 237 |
| Peak RSS, serving (MiB) | 66 | 74 | **26** | 438 | 38 | 141 | 84 |
| On-disk size (MiB) | 9.8 | **9.8** | 63.0 | 10.3 | 13.7 | 48.6 | 14.2 |
| Cold start to first answer (s) | 0.62 | 0.63 | 0.09 | 0.45 | **0.07** | 0.29 | 0.24 |
| Replace one document (durable) (ms) | 3.919 | **3.805** | 4.626 | 18.039 | 163.897 | 146.631 | n/m (immutable index; any change is a full rebuild) |
| Delete one document (durable) (ms) | 3.162 | 3.269 | 6.418 | **1.482** | 176.531 | 16.973 | n/m (immutable index; any change is a full rebuild) |
| Install size (MiB) | 2.4 | 2.4 | **0.0** | 5.7 | 13.8 | 2.0 | 56.5 |
| Install distributions (count) | 92 | 92 | **0** | 2 | 1 | 2 | 3 |
Bold marks the best value in the row.
</details>
### memory (4,559 documents, 100 queries)
Suite `standard`, hardware class `linux-x86_64-intel-ultra-9-185h-22c-27g` (Intel(R) Core(TM) Ultra 9 185H, 22 logical cores, 27 GiB, Linux 6.18.33.2-microsoft-standard-WSL2), commit `unknown`, 2026-10-21. Raw results: [standard-text-memory-linux-x86_64-intel-ultra-9-185h-22c-27g.json](../../benchmarks/results/competitive/standard-text-memory-linux-x86_64-intel-ultra-9-185h-22c-27g.json). Intervals: 95.00% bootstrap (Bonferroni family of 0 verdict metrics).
Timing boundary (where each query is timed): WARM: native Rust API call in the engine process (no language binding); text: BM25 index only (`--encode lexical`); ANN: public-bench loop; SQLite FTS5: CPython sqlite3 call per query (query string built outside the timer); Apache Lucene 9.12.3 (JVM): in-process JVM API call, after a full untimed JIT warm-up pass; tantivy: CPython tantivy-py call per query (stored id fetched per hit).
| Metric | WARM | Best other | Advantage (WARM vs best other) | Verdict |
|---|---|---|---|---|
| nDCG@10 | 0.4561† | Apache Lucene 9.12.3 (JVM) 0.4627† | n/v (no CI) | **n/v** |
| recall@100 | 0.8203† | SQLite FTS5 0.8392† | n/v (no CI) | **n/v** |
| QPS (single thread) (QPS) | 12,629† | tantivy 4,850† | n/v (no CI) | **n/v** |
| p50 latency (us) | 76† | tantivy 190† | n/v (no CI) | **n/v** |
| p99 latency (us) | 145† | tantivy 396† | n/v (no CI) | **n/v** |
| Ingest/build time (s) | 1.21† | SQLite FTS5 0.88† | n/v (no CI) | **n/v** |
| Peak RSS, build (MiB) | n/m (Linux child inherited the orchestrator high-water mark; measured before the isolated-spawner fix) | n/m | n/m | **n/m** |
| Peak RSS, serving (MiB) | n/m (Linux child inherited the orchestrator high-water mark; measured before the isolated-spawner fix) | n/m | n/m | **n/m** |
| On-disk size (MiB) | 9.8† | Apache Lucene 9.12.3 (JVM) 10.3† | n/v (no CI) | **n/v** |
| Cold start to first answer (s) | 0.30† | tantivy 0.02† | n/v (no CI) | **n/v** |
| Replace one document (durable) (ms) | 4.457† | SQLite FTS5 11.853† | n/v (no CI) | **n/v** |
| Delete one document (durable) (ms) | 3.331† | SQLite FTS5 10.456† | n/v (no CI) | **n/v** |
| Install size (MiB) | 2.5† | SQLite FTS5 0.0† | n/v (no CI) | **n/v** |
| Install distributions (count) | 92† | SQLite FTS5 0† | n/v (no CI) | **n/v** |
<details><summary>All systems</summary>
| Metric | WARM | SQLite FTS5 | Apache Lucene 9.12.3 (JVM) | tantivy |
|---|---|---|---|---|
| nDCG@10 | 0.4561 | 0.4501† | 0.4627† | 0.4491† |
| recall@100 | 0.8203 | 0.8392† | 0.8237† | 0.8203† |
| QPS (single thread) (QPS) | 12,629 | 63† | 1,338† | 4,850† |
| p50 latency (us) | 76 | 12,432† | 689† | 190† |
| p99 latency (us) | 145 | 61,157† | 1,366† | 396† |
| Ingest/build time (s) | 1.21 | 0.88† | 2.26† | 1.74† |
| Peak RSS, build (MiB) | n/m (Linux child inherited the orchestrator high-water mark; measured before the isolated-spawner fix) | n/m (Linux child inherited the orchestrator high-water mark; measured before the isolated-spawner fix) | n/m (Linux child inherited the orchestrator high-water mark; measured before the isolated-spawner fix) | n/m (Linux child inherited the orchestrator high-water mark; measured before the isolated-spawner fix) |
| Peak RSS, serving (MiB) | n/m (Linux child inherited the orchestrator high-water mark; measured before the isolated-spawner fix) | n/m (Linux child inherited the orchestrator high-water mark; measured before the isolated-spawner fix) | n/m (Linux child inherited the orchestrator high-water mark; measured before the isolated-spawner fix) | n/m (Linux child inherited the orchestrator high-water mark; measured before the isolated-spawner fix) |
| On-disk size (MiB) | 9.8 | 19.3† | 10.3† | 13.7† |
| Cold start to first answer (s) | 0.30 | 0.04† | 0.28† | 0.02† |
| Replace one document (durable) (ms) | 4.457 | 11.853† | 38.057† | 93.667† |
| Delete one document (durable) (ms) | 3.331 | 10.456† | 17.639† | 52.884† |
| Install size (MiB) | 2.5 | 0.0† | 5.7† | 13.8† |
| Install distributions (count) | 92 | 0† | 2† | 1† |
Bold marks the best value in the row.
</details>
† WARM vs Apache Lucene 9.12.3 (JVM): source revision is not pinned; no verdict.
† WARM vs SQLite FTS5: source revision is not pinned; no verdict.
† WARM vs tantivy: source revision is not pinned; no verdict.
### beir:nfcorpus (3,633 documents, 324 queries)
Suite `standard`, hardware class `linux-x86_64-intel-ultra-9-185h-22c-27g` (Intel(R) Core(TM) Ultra 9 185H, 22 logical cores, 27 GiB, Linux 6.18.33.2-microsoft-standard-WSL2), commit `unknown`, 2026-10-21. Raw results: [standard-text-beir-nfcorpus-linux-x86_64-intel-ultra-9-185h-22c-27g.json](../../benchmarks/results/competitive/standard-text-beir-nfcorpus-linux-x86_64-intel-ultra-9-185h-22c-27g.json). Intervals: 95.00% bootstrap (Bonferroni family of 0 verdict metrics).
Timing boundary (where each query is timed): WARM: native Rust API call in the engine process (no language binding); text: BM25 index only (`--encode lexical`); ANN: public-bench loop; SQLite FTS5: CPython sqlite3 call per query (query string built outside the timer); Apache Lucene 9.12.3 (JVM): in-process JVM API call, after a full untimed JIT warm-up pass; tantivy: CPython tantivy-py call per query (stored id fetched per hit).
| Metric | WARM | Best other | Advantage (WARM vs best other) | Verdict |
|---|---|---|---|---|
| nDCG@10 | 0.2916† | SQLite FTS5 0.2914† | n/v (no CI) | **n/v** |
| recall@100 | 0.2274† | Apache Lucene 9.12.3 (JVM) 0.2276† | n/v (no CI) | **n/v** |
| QPS (single thread) (QPS) | 40,784† | tantivy 15,413† | n/v (no CI) | **n/v** |
| p50 latency (us) | 26† | tantivy 71† | n/v (no CI) | **n/v** |
| p99 latency (us) | 57† | tantivy 150† | n/v (no CI) | **n/v** |
| Ingest/build time (s) | 0.23† | SQLite FTS5 0.14† | n/v (no CI) | **n/v** |
| Peak RSS, build (MiB) | n/m (Linux child inherited the orchestrator high-water mark; measured before the isolated-spawner fix) | n/m | n/m | **n/m** |
| Peak RSS, serving (MiB) | n/m (Linux child inherited the orchestrator high-water mark; measured before the isolated-spawner fix) | n/m | n/m | **n/m** |
| On-disk size (MiB) | 2.1† | Apache Lucene 9.12.3 (JVM) 1.5† | n/v (no CI) | **n/v** |
| Cold start to first answer (s) | 0.09† | tantivy 0.02† | n/v (no CI) | **n/v** |
| Replace one document (durable) (ms) | 2.033† | SQLite FTS5 9.738† | n/v (no CI) | **n/v** |
| Delete one document (durable) (ms) | 1.714† | SQLite FTS5 9.432† | n/v (no CI) | **n/v** |
| Install size (MiB) | 2.5† | SQLite FTS5 0.0† | n/v (no CI) | **n/v** |
| Install distributions (count) | 92† | SQLite FTS5 0† | n/v (no CI) | **n/v** |
<details><summary>All systems</summary>
| Metric | WARM | SQLite FTS5 | Apache Lucene 9.12.3 (JVM) | tantivy |
|---|---|---|---|---|
| nDCG@10 | 0.2916 | 0.2914† | 0.2897† | 0.2901† |
| recall@100 | 0.2274 | 0.2271† | 0.2276† | 0.2269† |
| QPS (single thread) (QPS) | 40,784 | 1,416† | 4,115† | 15,413† |
| p50 latency (us) | 26 | 255† | 283† | 71† |
| p99 latency (us) | 57 | 2,701† | 539† | 150† |
| Ingest/build time (s) | 0.23 | 0.14† | 0.62† | 0.28† |
| Peak RSS, build (MiB) | n/m (Linux child inherited the orchestrator high-water mark; measured before the isolated-spawner fix) | n/m (Linux child inherited the orchestrator high-water mark; measured before the isolated-spawner fix) | n/m (Linux child inherited the orchestrator high-water mark; measured before the isolated-spawner fix) | n/m (Linux child inherited the orchestrator high-water mark; measured before the isolated-spawner fix) |
| Peak RSS, serving (MiB) | n/m (Linux child inherited the orchestrator high-water mark; measured before the isolated-spawner fix) | n/m (Linux child inherited the orchestrator high-water mark; measured before the isolated-spawner fix) | n/m (Linux child inherited the orchestrator high-water mark; measured before the isolated-spawner fix) | n/m (Linux child inherited the orchestrator high-water mark; measured before the isolated-spawner fix) |
| On-disk size (MiB) | 2.1 | 4.0† | 1.5† | 2.0† |
| Cold start to first answer (s) | 0.09 | 0.02† | 0.28† | 0.02† |
| Replace one document (durable) (ms) | 2.033 | 9.738† | 30.876† | 84.138† |
| Delete one document (durable) (ms) | 1.714 | 9.432† | 14.463† | 42.872† |
| Install size (MiB) | 2.5 | 0.0† | 5.7† | 13.8† |
| Install distributions (count) | 92 | 0† | 2† | 1† |
Bold marks the best value in the row.
</details>
† WARM vs SQLite FTS5: source revision is not pinned; no verdict.
† WARM vs Apache Lucene 9.12.3 (JVM): source revision is not pinned; no verdict.
† WARM vs tantivy: source revision is not pinned; no verdict.
### beir:nfcorpus (3,633 documents, 324 queries)
Suite `heavy`, hardware class `modal-x86_64-8vcpu-32g` (unknown, 24 logical cores, 396 GiB, Linux 4.19.0-gvisor), commit `d5f3f3b8cb`, 2026-10-21. Raw results: [heavy-text-beir-nfcorpus-modal-x86_64-8vcpu-32g.json](../../benchmarks/results/competitive/heavy-text-beir-nfcorpus-modal-x86_64-8vcpu-32g.json). Intervals: 99.50% bootstrap (Bonferroni family of 10 verdict metrics).
Timing boundary (where each query is timed): WARM: native Rust API call in the engine process (no language binding); text: BM25 index only (`--encode lexical`); ANN: public-bench loop; WARM (eager hypervectors): native Rust API call in the engine process (no language binding); text: BM25 index plus chunk hypervectors at ingest (`--encode eager`); SQLite FTS5: CPython sqlite3 call per query (query string built outside the timer); Apache Lucene 9.12.3 (JVM): in-process JVM API call, after a full untimed JIT warm-up pass; tantivy: CPython tantivy-py call per query (stored id fetched per hit); Whoosh: pure Python; bm25s: CPython + NumPy (single-threaded retrieve).
| Metric | WARM | Best other | Advantage (WARM vs best other) | Verdict |
|---|---|---|---|---|
| nDCG@10 | 0.2916 | SQLite FTS5 0.2914 | 1.00x [0.70, 1.42] | **TIE** |
| recall@100 | 0.2274 | Apache Lucene 9.12.3 (JVM) 0.2276 | 1.00x [0.69, 1.45] | **TIE** |
| QPS (single thread) (QPS) | 10,572 | tantivy 7,151 | 1.48x [1.17, 1.63] | **WIN** |
| p50 latency (us) | 60 | tantivy 157 | 2.64x [2.00, 2.93] | **WIN** |
| p99 latency (us) | 279 | tantivy 294 | 1.05x [0.84, 1.21] | **TIE** |
| Ingest/build time (s) | 0.59 | SQLite FTS5 0.29 | 0.48x [0.47, 0.49] | **LOSS** |
| Peak RSS, build (MiB) | 57 | SQLite FTS5 34 | 0.59x [0.58, 0.60] | **LOSS** |
| Peak RSS, serving (MiB) | 57 | SQLite FTS5 30 | 0.53x [0.50, 0.54] | **LOSS** |
| On-disk size (MiB) | 2.1 | Apache Lucene 9.12.3 (JVM) 1.5 | 0.72x (no CI) | **LOSS** |
| Cold start to first answer (s) | 0.18 | tantivy 0.07 | 0.39x [0.35, 0.43] | **LOSS** |
| Replace one document (durable) (ms) | 1.073 | SQLite FTS5 3.141 | 2.93x (no CI) | **n/v** |
| Delete one document (durable) (ms) | 0.729 | Apache Lucene 9.12.3 (JVM) 1.349 | 1.85x (no CI) | **n/v** |
| Install size (MiB) | 2.4 | Whoosh 2.0 | 0.83x (no CI) | **n/v** |
| Install distributions (count) | 92 | tantivy 1 | 0.01x (no CI) | **n/v** |
<details><summary>All systems</summary>
| Metric | WARM | WARM (eager hypervectors) | SQLite FTS5 | Apache Lucene 9.12.3 (JVM) | tantivy | Whoosh | bm25s |
|---|---|---|---|---|---|---|---|
| nDCG@10 | **0.2916** | 0.2916 | 0.2914 | 0.2897 | 0.2901 | 0.2873 | 0.2878 |
| recall@100 | 0.2274 | 0.2274 | 0.2271 | **0.2276** | 0.2269 | 0.2252 | 0.2272 |
| QPS (single thread) (QPS) | 10,572 | **10,809** | 486 | 3,023 | 7,151 | 131 | 1,233 |
| p50 latency (us) | 60 | **57** | 775 | 406 | 157 | 2,969 | 799 |
| p99 latency (us) | 279 | **275** | 6,963 | 703 | 294 | 33,136 | 1,082 |
| Ingest/build time (s) | 0.59 | 2.18 | **0.29** | 0.97 | 0.49 | 7.33 | 0.88 |
| Peak RSS, build (MiB) | 57 | 87 | **34** | 455 | 78 | 144 | 92 |
| Peak RSS, serving (MiB) | 57 | 58 | **30** | 441 | 41 | 70 | 70 |
| On-disk size (MiB) | 2.1 | 2.1 | 9.3 | **1.5** | 2.0 | 9.9 | 3.4 |
| Cold start to first answer (s) | 0.18 | 0.18 | 0.07 | 0.44 | **0.07** | 0.24 | 0.21 |
| Replace one document (durable) (ms) | 1.073 | **1.065** | 3.141 | 18.216 | 55.324 | 53.212 | n/m (immutable index; any change is a full rebuild) |
| Delete one document (durable) (ms) | 0.729 | **0.706** | 8.833 | 1.349 | 50.475 | 6.141 | n/m (immutable index; any change is a full rebuild) |
| Install size (MiB) | 2.4 | 2.4 | **0.0** | 5.7 | 13.8 | 2.0 | 56.5 |
| Install distributions (count) | 92 | 92 | **0** | 2 | 1 | 2 | 3 |
Bold marks the best value in the row.
</details>
### beir:scifact (5,183 documents, 809 queries)
Suite `standard`, hardware class `linux-x86_64-intel-ultra-9-185h-22c-27g` (Intel(R) Core(TM) Ultra 9 185H, 22 logical cores, 27 GiB, Linux 6.18.33.2-microsoft-standard-WSL2), commit `unknown`, 2026-10-21. Raw results: [standard-text-beir-scifact-linux-x86_64-intel-ultra-9-185h-22c-27g.json](../../benchmarks/results/competitive/standard-text-beir-scifact-linux-x86_64-intel-ultra-9-185h-22c-27g.json). Intervals: 95.00% bootstrap (Bonferroni family of 0 verdict metrics).
Timing boundary (where each query is timed): WARM: native Rust API call in the engine process (no language binding); text: BM25 index only (`--encode lexical`); ANN: public-bench loop; SQLite FTS5: CPython sqlite3 call per query (query string built outside the timer); Apache Lucene 9.12.3 (JVM): in-process JVM API call, after a full untimed JIT warm-up pass; tantivy: CPython tantivy-py call per query (stored id fetched per hit).
| Metric | WARM | Best other | Advantage (WARM vs best other) | Verdict |
|---|---|---|---|---|
| nDCG@10 | 0.7027† | Apache Lucene 9.12.3 (JVM) 0.7022† | n/v (no CI) | **n/v** |
| recall@100 | 0.9264† | SQLite FTS5 0.9322† | n/v (no CI) | **n/v** |
| QPS (single thread) (QPS) | 17,782† | tantivy 6,450† | n/v (no CI) | **n/v** |
| p50 latency (us) | 55† | tantivy 151† | n/v (no CI) | **n/v** |
| p99 latency (us) | 95† | tantivy 287† | n/v (no CI) | **n/v** |
| Ingest/build time (s) | 0.31† | SQLite FTS5 0.20† | n/v (no CI) | **n/v** |
| Peak RSS, build (MiB) | 73† | SQLite FTS5 73† | n/v (no CI) | **n/v** |
| Peak RSS, serving (MiB) | 196† | tantivy 174† | n/v (no CI) | **n/v** |
| On-disk size (MiB) | 2.9† | Apache Lucene 9.12.3 (JVM) 2.0† | n/v (no CI) | **n/v** |
| Cold start to first answer (s) | 0.12† | tantivy 0.02† | n/v (no CI) | **n/v** |
| Replace one document (durable) (ms) | 2.004† | SQLite FTS5 9.132† | n/v (no CI) | **n/v** |
| Delete one document (durable) (ms) | 1.730† | SQLite FTS5 8.999† | n/v (no CI) | **n/v** |
| Install size (MiB) | 2.5† | SQLite FTS5 0.0† | n/v (no CI) | **n/v** |
| Install distributions (count) | 92† | SQLite FTS5 0† | n/v (no CI) | **n/v** |
<details><summary>All systems</summary>
| Metric | WARM | SQLite FTS5 | Apache Lucene 9.12.3 (JVM) | tantivy |
|---|---|---|---|---|
| nDCG@10 | 0.7027 | 0.7009† | 0.7022† | 0.7013† |
| recall@100 | 0.9264 | 0.9322† | 0.9295† | 0.9283† |
| QPS (single thread) (QPS) | 17,782 | 242† | 2,862† | 6,450† |
| p50 latency (us) | 55 | 3,853† | 329† | 151† |
| p99 latency (us) | 95 | 9,343† | 789† | 287† |
| Ingest/build time (s) | 0.31 | 0.20† | 0.71† | 0.37† |
| Peak RSS, build (MiB) | 73 | 73† | 162† | 73† |
| Peak RSS, serving (MiB) | 196 | 218† | 410† | 174† |
| On-disk size (MiB) | 2.9 | 5.9† | 2.0† | 2.6† |
| Cold start to first answer (s) | 0.12 | 0.03† | 0.29† | 0.02† |
| Replace one document (durable) (ms) | 2.004 | 9.132† | 30.774† | 86.064† |
| Delete one document (durable) (ms) | 1.730 | 8.999† | 14.387† | 44.826† |
| Install size (MiB) | 2.5 | 0.0† | 5.7† | 13.8† |
| Install distributions (count) | 92 | 0† | 2† | 1† |
Bold marks the best value in the row.
</details>
† WARM vs Apache Lucene 9.12.3 (JVM): source revision is not pinned; no verdict.
† WARM vs SQLite FTS5: source revision is not pinned; no verdict.
† WARM vs tantivy: source revision is not pinned; no verdict.
### beir:scifact (5,183 documents, 809 queries)
Suite `heavy`, hardware class `modal-x86_64-8vcpu-32g` (unknown, 24 logical cores, 396 GiB, Linux 4.19.0-gvisor), commit `d5f3f3b8cb`, 2026-10-21. Raw results: [heavy-text-beir-scifact-modal-x86_64-8vcpu-32g.json](../../benchmarks/results/competitive/heavy-text-beir-scifact-modal-x86_64-8vcpu-32g.json). Intervals: 99.50% bootstrap (Bonferroni family of 10 verdict metrics).
Timing boundary (where each query is timed): WARM: native Rust API call in the engine process (no language binding); text: BM25 index only (`--encode lexical`); ANN: public-bench loop; WARM (eager hypervectors): native Rust API call in the engine process (no language binding); text: BM25 index plus chunk hypervectors at ingest (`--encode eager`); SQLite FTS5: CPython sqlite3 call per query (query string built outside the timer); Apache Lucene 9.12.3 (JVM): in-process JVM API call, after a full untimed JIT warm-up pass; tantivy: CPython tantivy-py call per query (stored id fetched per hit); Whoosh: pure Python; bm25s: CPython + NumPy (single-threaded retrieve).
| Metric | WARM | Best other | Advantage (WARM vs best other) | Verdict |
|---|---|---|---|---|
| nDCG@10 | 0.7027 | Apache Lucene 9.12.3 (JVM) 0.7022 | 1.00x [0.90, 1.11] | **TIE** |
| recall@100 | 0.9264 | SQLite FTS5 0.9322 | 0.99x [0.94, 1.05] | **TIE** |
| QPS (single thread) (QPS) | 2,733 | tantivy 2,965 | 0.92x [0.87, 0.98] | **LOSS** |
| p50 latency (us) | 373 | tantivy 334 | 0.90x [0.85, 0.96] | **LOSS** |
| p99 latency (us) | 518 | tantivy 508 | 0.98x [0.87, 1.19] | **TIE** |
| Ingest/build time (s) | 1.42 | SQLite FTS5 0.43 | 0.30x [0.28, 0.32] | **LOSS** |
| Peak RSS, build (MiB) | 65 | SQLite FTS5 39 | 0.60x [0.59, 0.61] | **LOSS** |
| Peak RSS, serving (MiB) | 160 | SQLite FTS5 54 | 0.34x [0.34, 0.35] | **LOSS** |
| On-disk size (MiB) | 2.9 | Apache Lucene 9.12.3 (JVM) 2.0 | 0.68x (no CI) | **LOSS** |
| Cold start to first answer (s) | 0.24 | SQLite FTS5 0.08 | 0.32x [0.27, 0.37] | **LOSS** |
| Replace one document (durable) (ms) | 1.016 | SQLite FTS5 3.199 | 3.15x (no CI) | **n/v** |
| Delete one document (durable) (ms) | 0.793 | Apache Lucene 9.12.3 (JVM) 1.269 | 1.60x (no CI) | **n/v** |
| Install size (MiB) | 2.4 | Whoosh 2.0 | 0.83x (no CI) | **n/v** |
| Install distributions (count) | 92 | tantivy 1 | 0.01x (no CI) | **n/v** |
<details><summary>All systems</summary>
| Metric | WARM | WARM (eager hypervectors) | SQLite FTS5 | Apache Lucene 9.12.3 (JVM) | tantivy | Whoosh | bm25s |
|---|---|---|---|---|---|---|---|
| nDCG@10 | **0.7027** | 0.7027 | 0.7009 | 0.7022 | 0.7013 | 0.6908 | 0.7001 |
| recall@100 | 0.9264 | 0.9264 | **0.9322** | 0.9295 | 0.9283 | 0.9186 | 0.9283 |
| QPS (single thread) (QPS) | 2,733 | 2,702 | 99 | 1,457 | **2,965** | 23 | 1,127 |
| p50 latency (us) | 373 | 376 | 9,931 | 642 | **334** | 36,023 | 871 |
| p99 latency (us) | 518 | 530 | 17,709 | 1,128 | **508** | 128,595 | 1,259 |
| Ingest/build time (s) | 1.42 | 3.60 | **0.43** | 1.52 | 0.76 | 11.38 | 1.40 |
| Peak RSS, build (MiB) | 65 | 96 | **39** | 504 | 90 | 178 | 104 |
| Peak RSS, serving (MiB) | 160 | 160 | **54** | 952 | 65 | 101 | 95 |
| On-disk size (MiB) | 2.9 | 2.9 | 12.4 | **2.0** | 2.6 | 13.2 | 4.5 |
| Cold start to first answer (s) | 0.24 | 0.25 | **0.08** | 0.46 | 0.08 | 0.24 | 0.23 |
| Replace one document (durable) (ms) | 1.016 | **0.923** | 3.199 | 16.979 | 73.369 | 48.782 | n/m (immutable index; any change is a full rebuild) |
| Delete one document (durable) (ms) | 0.793 | **0.714** | 6.280 | 1.269 | 63.293 | 6.583 | n/m (immutable index; any change is a full rebuild) |
| Install size (MiB) | 2.4 | 2.4 | **0.0** | 5.7 | 13.8 | 2.0 | 56.5 |
| Install distributions (count) | 92 | 92 | **0** | 2 | 1 | 2 | 3 |
Bold marks the best value in the row.
</details>
## Operational facts
| System | Service needed | Encrypted at rest | Basis |
|---|---|---|---|
| hnswlib | no | no | documentation (not exercised) |
| WARM | no | yes | measured: benchmarks/results/security_matrix_v2.json |
| SQLite FTS5 | no | no (SQLCipher is a separate build) | documentation (not exercised) |
| Apache Lucene 9.12.3 (JVM) | no (library; needs a JVM) | no (filesystem-level only) | documentation (not exercised) |
| tantivy | no | no | documentation (not exercised) |
| WARM (eager hypervectors) | no | yes | measured: benchmarks/results/security_matrix_v2.json |
| Whoosh | no | no | documentation (not exercised) |
| bm25s | no | no | documentation (not exercised) |
## Versions
| System | Pinned version(s) |
|---|---|
| hnswlib | hnswlib 0.8.0 |
| WARM | holographic-memory 0.6.1, commit d5f3f3b8cbf963003d282181409e6a0203f9e96d |
| SQLite FTS5 | python-sqlite3 3.40.1 |
| Apache Lucene 9.12.3 (JVM) | lucene 9.12.3, jvm openjdk version "17.0.20.1" 2026-08-18 |
| tantivy | tantivy 0.26.2 |
| WARM (eager hypervectors) | holographic-memory 0.6.1, commit d5f3f3b8cbf963003d282181409e6a0203f9e96d |
| Whoosh | whoosh-reloaded 2.7.5 |
| bm25s | bm25s 0.3.14, PyStemmer 3.1.0 |
## Not run
- `pgvector`: needs a PostgreSQL server with the extension; not installable from pip/cargo/npm (service required)
- `mem0`: agent-memory systems belong to the next harness stage (plug in through the same adapter interface)
- `zep`: agent-memory systems belong to the next harness stage
- `letta`: agent-memory systems belong to the next harness stage
- `langmem`: agent-memory systems belong to the next harness stage