vemb

httpie for embeddings. Embed text, images, audio, video, and PDFs from the command line.

pipx install vemb
export GEMINI_API_KEY=your_key
vemb text "hello world"

Powered by Gemini Embedding 2, the first natively multimodal embedding model. One model, one vector space for everything.

Install

pipx install vemb
# or
pip install vemb

Get a free API key at https://aistudio.google.com/apikey

export GEMINI_API_KEY=your_key

Commands

vemb text "hello world"                    # embed text
vemb embed photo.jpg                       # embed any file (auto-detects type)
vemb embed *.jpg --jsonl                   # batch embed, one JSON per line
vemb image photo.jpg                       # embed image (PNG, JPEG)
vemb audio clip.mp3                        # embed audio (MP3, WAV)
vemb video clip.mp4                        # embed video (MP4, MOV)
vemb pdf doc.pdf                           # embed PDF
vemb similar photo1.jpg photo2.jpg         # cosine similarity between two files
vemb search ./photos "sunset at beach"     # search a directory

Pipe from stdin:

echo "hello world" | vemb text -
cat document.txt | vemb text -

Output

Default output is JSON:

{
  "model": "gemini-embedding-2-preview",
  "dimensions": 768,
  "values": [0.012, -0.034, ...]
}

Options:

vemb text "hello" --compact                # just the vector array
vemb text "hello" --numpy                  # numpy format
vemb text "hello" --dim 768                # set dimensions (128-3072)
vemb text "hello" --task RETRIEVAL_QUERY   # set task type

Batch mode outputs JSONL (one embedding per line):

vemb embed *.jpg --jsonl > embeddings.jsonl

Search

Search indexes a directory and finds files similar to your query:

vemb search ./photos "sunset at beach" --top 5

Embeddings are cached as a binary numpy matrix in .vemb/vectors.npy with a lightweight .vemb/manifest.json mapping keys to row indices. Vectors are pre-normalized so cosine reduces to a dot product at query time. Unchanged files won't be re-embedded on subsequent searches.

Legacy .vemb/cache.json caches (from v0.2.0 and earlier) are migrated to the binary format automatically on first load.

Supported formats

Type	Formats
Text	any string, stdin
Image	PNG, JPEG
Audio	MP3, WAV (up to 80s)
Video	MP4, MOV (up to 128s)
PDF	up to 6 pages

License

MIT