Static Embeddings 04 EN

Static semantic embeddings 04: Go parity and packed execution, weak code relevance

한국어 · Frozen plan · Results · Source hashes

A pure-Go static embedding runner now matches upstream and keeps ternary weights packed in memory. CoSQA validation quality remains below our lexical control, so no promotion. Reserve1 (2,403) and reserve2 (2,401) remain unscored.

Implementation and provenance

MinishLab Model2Vec's potion-base-2M is a general-English static embedding whose model card declares MIT. Pin model revision 389b9f64be5aa4ae7a6bc6fe95ef20ce485ae5da and reference implementation commit 38c7bc9c93808ee15602265ba96a9b8fda58783f. This is a separate comparison family, not a port/replacement of Laya's encoder.

Go implements WordPiece/BertNormalizer, averages64-dimensional token vectors and normalizes. No automatic CLS/SEP; unknown IDs are removed as upstream does, while literal special tokens are handled separately. Empty or all-unknown input yields a zero vector, not meaningful relevance evidence. Inputs must be valid UTF-8 and at most4,096 bytes; oversized inputs fail instead of silently truncating.

The existing Laya NFC/ByteLevel/BPE tokenizer is incompatible and is not reused. Exact weight/tokenizer hashes are mandatory. Runtime vocabulary uses sorted arrays and weights are contiguous; a temporary map is used during JSON loading only. Inference has no maps, mutable cache or locks; calls own scratch/results.

Reference checks

• 34 newly authored public sentences cover accents, Korean/Chinese/Japanese, literal specials, controls, emoji, empty input and long words. All token IDs match; max FP32 absolute error8.94e-8 versus tolerance2e-6.
• 1,214 texts from all607 validation query/code pairs also match IDs exactly; max error1.19e-7. This tokenizer check includes the54 pairs outside the earlier lexical scope. Relevance AUC still uses the same553 eligible pairs. No reserve is used.
• Pinned Python Model2Vec is a maintainer oracle only. User execution, downloading and actual CI tests require no Python/ONNX/Torch. Only34 original public reference cases enter Git; CoSQA source/vector traces remain local.
• Race tests, CGO-disabled execution, explicit packed arithmetic examples, corrupt assets and input-budget failures are covered. CI downloads verified real assets and checks saved reference outputs, not mocks alone.

Compression and measured cost

Mode	In-memory weights plus scales	Validation AUC	Replay short-text encode p95
FP32	7,559,168B	0.5289	5.04µs
INT8, per-token scale	2,007,904B	0.5286	3.75µs
Ternary, per-token scale	590,560B (576.72KiB)	0.5394	4.83µs
BM25 control	separate	0.5378	not measured here
Previous16-coefficient matching FP32	separate	0.5729	not measured here

These are point estimates on the same553 eligible validation pairs, not confidence intervals or fresh final results. Report54 exclusions explicitly. Primary FP32 misses the prespecified AUC≥0.60 and ≥0.03 above BM25 conditions. Do not replace it post hoc with ternary because of a small apparent gain. No task-specific fitting occurred this cycle.

Ternary uses per-row mean absolute value as scale and retains signs at magnitude≥0.7×scale. Two bits/coefficient plus a four-byte scale per64 coefficients gives 2.5bits/coefficient, not1.58. Packed bits are read directly, but accumulation is float32; no GPU or specialized SIMD kernel. Startup reads the7.56MB FP32 source then packs it, so disk storage is not590KB.

Apple M4 Pro/24GiB/Go1.27.1,2,000 calls on one original public sentence. First p95 FP32/INT8/ternary:4.21/3.63/4.67µs; replay values above. Each912B/op and31allocs/op. Replay live Go heap after GC: approximately8.44/3.14/1.80MiB. First whole-process maxRSS including loading was31.73/33.58/32.33MiB: compressed modes did not lower this peak. Distinguish table size, live heap and process peak; tables exclude vocabulary/runtime. Short-text encode cost is not64-candidate retrieval or end-to-end agent latency.

Human/agent usage

# Once: download immutable public assets and verify SHA256
go run ./cmd/riido-staticprobe --stage setup
# Offline thereafter: one JSON object in, vector/token IDs/table bytes out
printf '%s\n' '{"text":"parse HTTP response headers"}' | \
  go run ./cmd/riido-staticprobe --mode ternary
# Requires the previously prepared pinned local CoSQA source
go run ./cmd/riido-staticprobe --stage validation --mode fp32 \
  --out .cache/static-validation-new.json

Validation output must be a new file. This is an experimental vector supplier; it changes neither the default riido-hints model nor Codex registration. Maintainer reference generator: scripts/training/static_reference.py; users and CI consume saved public references.

Next steps and licensing scope

Plain mean-vector cosine did not transfer general-English semantics to this code-relevance task. Retain the lexical control while testing identifier normalization, a small learned query/code alignment layer and, if justified, task-specific static embedding tuning. Extra memory without quality gains is not a reason to adopt a model. Go packed execution is now available, but candidate retrieval and LLM savings remain unverified.

MIT is the model author's declaration, not proof of all upstream-data rights or absence of pretraining overlap. Preserve the CoSQA distinction between MIT code and C-UDA data; no raw rows/labels are redistributed. Upstream weights remain local and were not copied to Git or our HF account. New-model publication requires provenance, verification and license/content review.