v0.3.0: binary cache

What's new

Replace the per-key JSON cache with a binary numpy matrix + tiny JSON manifest. Vectors are pre-normalized on write so cosine reduces to a dot product at query time.

End-to-end benchmarks

At 3072-dim (Gemini Embedding 2), warm cache, best of 3 runs on Apple Silicon:

N	v0.2.0 (JSON + Python loop)	v0.3.0 (.npy + BLAS)	Speedup
200	1.62s	1.48s	tied
1000	2.34s	1.51s	1.5x
5000	4.31s	1.64s	2.6x

Cache size (N=5000, 3072-dim)

• v0.2.0: 317 MB cache.json
• v0.3.0: 61 MB vectors.npy + 229 KB manifest.json (5x smaller)

Other

• Cosine stays exact, no ANN, no approximation
• numpy is lazy-imported so vemb --version, vemb text, vemb embed etc stay fast
• Legacy .vemb/cache.json is auto-migrated on first load, no user action needed

Install

pip install -U vemb

Full diff: v0.2.0...v0.3.0

v0.2.0

• stdin piping: echo "text" | vemb text -
• batch embed: vemb embed *.jpg --jsonl
• fixed cache key collisions across subdirectories
• SDK error handling (clean messages instead of tracebacks)
• input validation for --dim and --top
• proper file/directory path constraints

https://pypi.org/project/vemb/0.2.0/

v0.1.0

Initial release.

pipx install vemb

• Embed text, images, audio, video, PDFs from the command line
• Cosine similarity between any two files
• Search a directory with caching
• Powered by Gemini Embedding 2

https://pypi.org/project/vemb/0.1.0/