Pancake
Wiki Knowledge Pack
All of Simple English Wikipedia in one searchable file. No search server, no embedding API, no data leaving this page — watch the meter.
This session
—
Last query
—
Every byte itemized. Encoding, search, and ranking
run in this tab; the network carries only file ranges.
booting…
What just happened
- Your words became a vector — MiniLM-L6 (45 MB, fp16) runs in this tab via WebAssembly. ~30 ms.
- A resident sketch of all 456,153 chunks was scanned — 47.4 MB of 4-bit sketches, held in memory, scanned with WASM SIMD. ~20 ms.
- The top ~200 candidates were fetched by byte range — ~66 HTTP range requests, ~390 KiB, against a plain static file.
- Exact rerank over the fetched rows picked the top 10.
- A calibrated scorer judged the match — strong, weak, or "this pack cannot help you," using retrieval signals plus a 256 KB vocabulary filter.
Measured, not promised
| Corpus | 241,787 articles → 456,153 chunks (Simple English Wikipedia, 2023-11) |
| Pack | 222.6 MB artifact · 47.4 MB resident tier, hash-verified at open |
| Recall@10 | 95.65% vs exact float search; 100% on hand-written questions |
| Per query | ~66 range requests · ~390 KiB · 0.6–1.5 s cold edge, ~330 ms warm edge, ~20 ms warm browser |
| Boot | ~93 MB one-time (encoder + resident tier), ~8 s cold (loads in parallel); ~2 s on repeat visits via service worker |
| Abstention | 14/14 golden probes pass in CI and in this page |
Honest limits
- First visit downloads ~93 MB before the first query (~8 s on a fast connection). Repeat visits boot in ~2 s from the service-worker cache — this design favors repeat use over drive-by page loads.
- The corpus is a 2023-11 snapshot. Ask about later events and the pack will confidently surface its best pre-event article.
- Abstention measures match strength, not intent: "best pizza near me" finds a real article about a pizza chain and calls it a match. It cannot know you meant delivery.
- The resident tier scales linearly (~104 bytes per chunk). This corpus costs 47 MB in your tab; full English Wikipedia would cost ~2.5 GB — bounded corpora are the point.