PipeOwl-1.8-jp-parameter-golf (Geometric Embedding)

A transformer-free semantic retrieval engine.

PipeOwl performs deterministic vocabulary scoring over a static embedding field:

score = α⋅base + β⋅Δfield

where:

• base = cosine similarity in embedding space
• Δfield = static scalar field bias

Features:

• O(n) over vocabulary.
• No attention.
• No transformer weights.
• CPU-friendly (<16MB model)

Architecture

• Static embedding table (V × D)
• Aligned vocabulary index
• Optional scalar bias field (Δfield)
• Linear scoring
• Pluggable decoder stage
• Targeted for CPU environments and low-latency systems (e.g. IME).

Model Specs

item	value
vocab size	26155
embedding dim	256
storage format	safetensors (FP16)
model size	~13.2 MB
languages	Japanese
startup time	<1s
query latency	34 ms (CPU, full vocabulary scan)

Quickstart

git clone https://huggingface.co/WangKaiLin/PipeOwl-1.8-jp-parameter-golf
cd PipeOwl-1.8-jp-parameter-golf

pip install numpy safetensors

python quickstart.py

Example:

Example semantic retrieval results:

Please enter words： 東京

Top-K Tokens:
1.000 | 東京
0.739 | 東京都
0.679 | 大阪
0.666 | ロンドン
0.646 | 名古屋

Please enter words： 大阪

Top-K Tokens:
1.000 | 大阪
0.756 | 関西
0.728 | 難波
0.717 | 京都
0.712 | 守口

Benchmark (CPU)

Environment:

• Vocab size: 26,155
• Embedding dimension: 256
• Hardware: CPU

Average query latency: maybe faster this is PipeOwl-1.6-jp benchmark

PipeOwl: 0.0036 sec BM25: 0.0421 sec Embedding: 0.0283 sec FAISS Flat: 0.0324 sec FAISS HNSW: 0.0230 sec

Comparison	Speedup
vs BM25	11.7× faster
vs Embedding	7.9× faster
vs FAISS Flat	9.0× faster
vs FAISS HNSW	6.4× faster

PipeOwl shows 6–12× lower latency compared with common retrieval baselines in this setup.

Repository Structure

PipeOwl-1.8-jp-parameter-golf/
 ├ README.md
 ├ config.json
 ├ LICENSE
 ├ quickstart.py
 ├ engine.py
 ├ vocabulary.json
 └ pipeowl_fp16.safetensors

LICENSE

MIT