PipeOwl-1.6-tw (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.

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	161783
embedding dim	1024
storage format	safetensors (FP16)
model size	~316 MB
languages	tw
startup time	~1s
query latency	~33-35 ms

Attribution

DATA_SOURCES.md

Quickstart

git clone https://huggingface.co/WangKaiLin/PipeOwl-1.6-tw
cd PipeOwl-1.6-tw

pip install numpy safetensors

python quickstart.py

Example:

Example semantic retrieval results:

請輸入句子： 今天

Top-K Tokens:
0.960 | 今天
0.893 | 今日
0.829 | 本日
0.827 | 今
0.822 | 這天

請輸入句子： 好睏

Top-K Tokens:
0.834 | 睏
0.834 | 好
0.792 | 好不
0.787 | 不錯
0.776 | 睏倦

請輸入句子： 水餃

Top-K Tokens:
0.962 | 水餃
0.831 | 餃
0.788 | 餃子
0.785 | 蒸餃
0.678 | 水漬

Repository Structure

pipeowl-1.6-tw/
 ├ README.md
 ├ config.json
 ├ DATA_SOURCES.md 
 ├ LICENSE
 ├ quickstart.py
 ├ engine.py
 ├ vocabulary.json
 └ pipeowl_fp16.safetensors

LICENSE

MIT