pikelet-wasm 0.11.0 → 0.12.0

package/CHANGELOG.md

@@ -8,6 +8,57 @@ Through 0.6.0 these were published as `pancake-wasm` and
 packages are pre-1.0: a minor bump may carry breaking changes, and each entry lists them
 first.
 
+## pikelet-wasm 0.12.0 / pikelet 0.12.0 — 2026-10-30
+
+Query embedding is about 1.65x faster, because the inline encoder now
+computes in integers, and `pikelet` no longer installs
+`@xenova/transformers`, which took a plain install from 82 packages and a
+critical audit advisory down to two packages. Upgrade if you embed queries
+in the browser or Node, or build `workers-ai` projects with `create`.
+Embeddings shift slightly (see Compatibility); packs compiled before this
+release still verify and serve. Both packages move to 0.12.0 together, and
+`pikelet` requires `pikelet-wasm` ^0.12.0.
+
+### Compatibility
+
+- **The inline encoder computes in integers, so its embeddings shift
+  slightly.** Every projection is now an integer GEMM: activations are
+  quantized to int16 per token and multiplied against the u8/u4 weights
+  with `i32x4.dot_i16x8_s`, and GELU is vectorized. Embeddings move by at
+  most about 3e-4 per component (MiniLM; 9e-5 for BGE-small), inside the
+  1e-3 tolerance packs declare for their test vectors, so every pack
+  compiled before this still verifies and serves: on every pack tested,
+  test vectors, golden queries and abstention probes gave the same results
+  under the old and new kernel. A pack rebuilt from the same inputs gets
+  slightly different vectors and so a new identity. The kernel uses only
+  baseline WASM SIMD, as before, and gives bit-identical embeddings in
+  Node, Chromium, Firefox and WebKit.
+
+- **`pikelet` no longer depends on `@xenova/transformers`.** `create`
+  (embedding mode `workers-ai`) now embeds passages at build time with the
+  inline-transformer kernel, running a u8 export of `BAAI/bge-small-en-v1.5`
+  rather than transformers.js's ONNX copy. The 36 MiB weight blob downloads
+  on the first build, pinned by sha256, into `~/.cache/pikelet/models`
+  (`PIKELET_MODEL_CACHE` moves it, `PIKELET_BGE_WEIGHTS_URL` names a
+  mirror); a cached blob that fails the pin is fetched again. Passage
+  vectors change slightly, so `pikelet rebuild` gives existing projects a
+  new index: they are within 0.9999 cosine of fp32 BGE-small where the old
+  path's were 0.996 to 0.998, and on BEIR SciFact, scored against live
+  Workers AI query vectors, nDCG@10 is 0.6991 against 0.6992 before.
+  Passages longer than 512 tokens are mean-pooled across windows instead of
+  truncated. `--omit=optional` is no longer needed: a plain install is two
+  packages, where it was 82 packages and six audit advisories (one critical)
+  inside that dependency.
+
+### Changed
+
+- **Query embedding is about 1.65x faster.** Single-query p50 on a Ryzen 9
+  4900HS: MiniLM 27.5 to 16.4 ms in Node (Chromium 26.4 to 15.8, WebKit
+  28.0 to 18.0), BGE-small u8 76.0 to 47.1 ms; long documents about 1.4x.
+  u4 weights gain little (about 1.06x): their 32-wide blocks each need a
+  conversion and scale. On BEIR SciFact with BGE-small, nDCG@10 is
+  unchanged (0.7071) and 298 of 300 top-10 lists are identical.
+
 ## pikelet-wasm 0.11.0 / pikelet 0.11.0 — 2026-10-24
 
 Pack reads on Node no longer go through the runtime's fetch, whose

package/README.md

@@ -44,7 +44,7 @@ console.log(out.matchQuality, out.results[0]?.title);
 // 'strong' 'Snapshot restore'
 ```
 
-`compile` also takes a live URL (`--source https://docs.example.com`), a repository's commit history (`--git ./repo`), or exported chats (`--conversations <file|folder>`). To install instead of using `npx`, run `npm install -g pikelet --omit=optional`; the optional dependency is only for `create`, which builds a deployed search app (a Worker and UI) rather than a file ([why it matters](packages/pikelet/README.md#installing)). [QUICKSTART.md](QUICKSTART.md) walks the whole lifecycle; [`packages/pikelet/README.md`](packages/pikelet/README.md) is the full CLI reference.
+`compile` also takes a live URL (`--source https://docs.example.com`), a repository's commit history (`--git ./repo`), or exported chats (`--conversations <file|folder>`). To install instead of using `npx`, run `npm install -g pikelet`. `create` builds a deployed search app (a Worker and UI) rather than a file. [QUICKSTART.md](QUICKSTART.md) walks the whole lifecycle; [`packages/pikelet/README.md`](packages/pikelet/README.md) is the full CLI reference.
 
 ## The shortest demonstration
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "pikelet-wasm",
-  "version": "0.11.0",
+  "version": "0.12.0",
   "description": "Vector search engine and .pikelet artifact readers/builders, in WebAssembly. Compile a corpus into one self-contained, hash-verified search file, or use the underlying HNSW engine directly. Runs in Node.js, browsers, and Cloudflare Workers with no native dependencies.",
   "main": "src/pikelet.js",
   "types": "src/pikelet.d.ts",