YellowPhoenix/fully-offline-semantic-search-apple-watch

Fully Offline Semantic Search on an Apple Watch and iPhone — 86.5% Top-10 @ 3.9M · 94% @ 4.9M · Airplane Mode

Measurement data kit + real-hardware demo videos for "Yellow Phoenix: Semantic Search on an Apple Watch — 86.5% Top-10 Fidelity at 3.9M Documents, Bit-Identical from watchOS to iOS to macOS and Linux" (Marc John Sawyer, 2026). Binary HNSW + 1024-bit re-rank; one continuous-take video shows the Watch answering in 34.8 ms, and the iPhone re-rank video was recorded in airplane mode — zero network, fully offline: 4.9M documents, 94.0% Top-10 @ ~0.2 ms on the device. Seven still frames verify every on-screen number.

Contents

• 512-bit ITQ binary hash index of ~5M academic papers (MiniLM embeddings)
• Binary HNSW graph structures
• 1024-bit re-rank payload sidecars
• Kill-test / three-judge fidelity measurement records

Dataset Viewer preview

The two preview_queries_*.jsonl configs load in the Dataset Viewer: every measured query (853 for 4.9M FAST, 695 for the 3.9M scan) with its corpus position and both hash representations — the exact bytes the devices searched. The multi-GB binary indexes stay in the tar for byte-identity with the Zenodo record.

Contents of yp_paper3_data.tar.gz

data/watch_ceiling/graph_1m_new512_m8_v5.bin      # 1M-node binary HNSW graph (v5 mmap)
data/watch_ceiling/flat_39m_new512.bin           # 3.9M-record 512-bit flat index (Tier C)
data/watch_ceiling/payload_sem_1m.json           # 1M semantic query payloads
data/watch_ceiling/payload_sem_39m.json          # 3.9M scan query payloads (695 q, h1k incl.)
data/watch_ceiling/payload_sem_49m.json          # 4.9M FAST query payloads (853 q, h1k incl.)
data/vines_10m/cluster_ids_bdc.u32               # BDC duplicate-cluster sidecar
benchmark_results/reconciliation_20261003.md     # reconciliation experiment (both conditionings)
benchmark_results/kill_test_ce_gt_20261003.json  # three-judge kill-test ground truth
benchmark_results/rerank_ksweep_20261003.log     # re-rank K-sweep (ef x K ceiling table)

Paper & provenance

• Cite the paper (concept DOI, always resolves to latest): https://doi.org/10.5281/zenodo.23118337
• This bundle is byte-identical to yp_paper3_data.tar.gz on the Zenodo record https://doi.org/10.5281/zenodo.23130414

Integrity

size	456,042,236 bytes
md5	677b276c63ecfff33f6d8e46290c76bd

Code

Engine source (Rust; macOS/iOS/watchOS/Linux; AGPL-3.0-or-later): https://github.com/Alfredai2025/yellow-phoenix-release

License

AGPL-3.0-or-later — free for research and open-source use; commercial licensing: marcsawyer76@outlook.com (see COMMERCIAL.md in the paper's source bundle).

Demo videos (real hardware, one continuous take each, audio removed)

File	Shows
iphone_semantic_4.9M_94_rerank_offline_airplane_mode.mp4	iPhone in airplane mode (fully offline): 4.9M FAST — graph 81.5–82.6% R@10 @ 176–792 µs, then 1024-bit re-rank lifts it to 94.0–94.6% R@10 — bit-identical to the Oct-2 runs (Nov 5 2026 re-run)
iphone_4.9M_94_rerank_airplane_mode_still.png	Full-resolution still of the offline 94% re-rank result screen (✈️ airplane mode visible in the status bar)
watch_semantic_3.9M_86.5_34.8ms.mp4	Watch: NEW 512 semantic 3.9M scan — 86.5% R@10 @ 34.8 ms on the wrist (the paper's headline run, one continuous take, Oct 2 2026)
iphone_semantic_3.9M_86.5_and_4.9M_FAST.mp4	iPhone: NEW 512 semantic 3.9M scan — 86.5% R@10 @ 7.5 ms (same fidelity as the Watch claim, ~4.6× faster) and 4.9M FAST — graph tier 82.6% R@10 @ 831 µs (the paper's 94% adds the 1024-bit re-rank)
watch_1m_pq_bench_with_mac_log.mp4	1M-record PQ bench on wrist with the synced Mac log on screen (dual-device proof)
watch_3m_flat_scan_100r10_27ms.mp4	3M-record flat/exhaustive scan: 100% R@10 at ~27 ms on the wrist; memory gate refusing the graph tier honestly (~574 MB needed vs 302 MB avail)
watch_106k_ism_full_sweep.mp4	Full 106K ISM sweep incl. 99.8 s on-watch graph build; HNSW R@10 99.7→100 %, p50 157 µs–1773 µs
watch_106k_ism_sweep_run2.mp4	Second run of the 106K sweep (repetition, numbers reproduce)
watch_table1_identical_1M_demo.mov	TABLE-1 bit-identity check at 1M (load 2.4 s, HNSW ef sweep, RAM recovered)
table1_still_01_menu.png … table1_still_07_ef400_memory_freed.png	Seven still frames from the TABLE-1 run — independent verification of every number in the demo video (load 2.4 s, memory gate OK, graph 0.50 s, ef ladder 501 µs → 1584 µs, RAM recovered)

Robustness & honest-failure demos (Sept 21 2026, perturbation stress-tests on the wrist)

File	Shows
watch_perturb_8b_full_run_flat_100r10_753us.mp4	Full perturbation run: FLAT scan keeps 100% R@10 at 753 µs even with corrupted queries, then HNSW perturb-8b ladder (99.7% R@1 @ 154 µs)
watch_flat_perturb_8b_24b_27ms_100r10.mp4	FLAT (Tier C) under perturb-8b/24b: 100% R@10 at ~27 ms — exhaustive scan is immune to query noise
watch_hnsw_perturb_48b_ef50_100r1_288us.mp4	HNSW with 48 of 512 bits flipped: still 100% R@1 @ 288 µs (ef=50); ef=25 already 99.7% @ 169 µs
watch_hnsw_perturb_24b_ladder_98r1.mp4	Perturb-24b ladder: 98.3% R@1 @ 667 µs (ef=100) — graceful degradation, no cliff
watch_hnsw_perturb_40b_ladder.mp4	Perturb-40b ladder: 93.3% R@1 @ 247 µs (ef=25) → 96.0% @ 303 µs (ef=50)
watch_hnsw_perturb_48b_ef400_ladder.mp4	Perturb-48b deep ladder: 93.3% @ 713 µs (ef=100) → 94.3% @ 1265 µs (ef=200), continuing to ef=400
watch_memory_gate_refused_574mb_needed.mp4	The honest refusal: memory gate says NO — "refused: need ~574 MB, avail 302 MB" — and falls back to exact search instead of crashing
watch_direct_exact_tierA_417ns_proof.mp4	Tier-A DIRECT exact index: R@1 290/300 at p50 417 ns, perturbed-miss 300/300 exact-match proof

Download

from huggingface_hub import hf_hub_download
p = hf_hub_download(repo_id="YellowPhoenix/fully-offline-semantic-search-apple-watch",
                    filename="yp_paper3_data.tar.gz")

Citation

@misc{sawyer2026yellowphoenixwatch,
  author = {Sawyer, Marc John},
  title  = {Yellow Phoenix: Semantic Search on an Apple Watch},
  year   = {2026},
  doi    = {10.5281/zenodo.23118337}
}