"Excuse me, I believe you have my stapler."
The in-process Node.js embeddings module for Flair. One precise thing — text → vector — done flawlessly, in-process, with no separate binary and no native runtime to install. Every embedding is verified against a pinned reference before it ships. The cover sheet that has to be right, or the whole report bounces.
Flair's memory is only as good as its embeddings, and today they come through a heavy inference runtime. Milton is the featherweight alternative: npm i and you have embeddings, in-process, no sidecar server, no GGUF-runtime dependency dragged into the install. It does the one job the overlooked, load-bearing worker always did — quietly, exactly, and without which nothing downstream works.
Milton is evidence-gated. Its output is defined as correct only when it matches a pinned reference implementation (nomic-embed-text-v1.5 via llama.cpp — the v1 oracle) within epsilon, over a conformance corpus — token-for-token discipline applied to vectors. An implementation that produces plausible-but-wrong vectors is the worst failure a memory layer can have: silent recall collapse, no error, nobody notices for weeks. So Milton fails closed on any unverified path — an embedding that hasn't cleared the golden-vector gate does not ship. (Reference discipline modeled on camelid.)
- GGUF-driven, nomic verified-first. Dequant and forward read architecture from the file — layer count, dims, pooling, normalization. We ship nomic-embed-text-v1.5 for v1 because Flair uses it today; that file is the golden-vector oracle, not a hardwired product. Prefix convention (
search_query:/search_document:) is config, not code. BERT-family (bge / gte / e5 / nomic-v2) is the same shape parameterized by the GGUF. Model #2 is pin goldens + prefix config + any GGUF-flagged arch variant (CLS vs mean), not a rewrite. - In-process, Node. No separate server, no OpenAI-compatible HTTP hop — a library you call.
- Next file, not a rewrite. A second BERT-family GGUF is goldens + prefix config + whatever pooling the file flags. No model registry.
Phase 1 (issue #1): the conformance + bench harness lives in harness/.
Tokenizer slice (issue #3): pure Rust nomic WordPiece + Flair prefixes in crate/.
GGUF load + dequant (issue #4): in crate/.
Forward pass (issue #5): native Rust nomic-bert — embeddings → layers → mean-pool → L2.
WASM-SIMD packaging (issue #6): same crate compiled wasm32 + simd128; src/
loads the prebuilt wasm/milton_bg.wasm (embed(text, {prefix}) -> Float32Array).
See harness/README.md, wasm/README.md, and docs/HARNESS-SPEC.md.
harness/ the golden-vector conformance + footprint/throughput bench
crate/ Rust tokenizer + GGUF dequant + forward (native bins + wasm32 lib)
src/ TS glue: embed(text, {prefix}) -> Float32Array, wasm loader
wasm/ prebuilt .wasm (SIMD) — not compiled at consumer install
docs/ HARNESS-SPEC.md, architecture, receipts