Skip to content
PITO MDdocs

Formats

pito-embed writes no files and prints no JSON. These are the shapes it promises to code that uses it.

Vectors

Section titled “Vectors”
  • embed_query returns a Vec<f32>; embed_documents returns a Vec<Vec<f32>>, one vector per input line, in the input’s order.
  • Every vector has embed::DIMS numbers: 768.
  • Every vector has unit length (L2 norm 1), so the dot product of two vectors is their cosine similarity, from -1 to 1.
  • Queries carry EmbeddingGemma’s query prompt (task: search result | query: ) and documents its document prompt (title: none | text: ), added by the library; compare queries with documents.
  • Each input is cut at 128 tokens.

The model folder

Section titled “The model folder”

Embedder::load(dir, device) reads six files from dir, laid out as on Hugging Face:

<dir>/
├── config.json
├── special_tokens_map.json
├── tokenizer.json
├── tokenizer_config.json
└── onnx/
    ├── model_q4.onnx
    └── model_q4.onnx_data

A missing file is EmbedError::ModelMissing with that file’s path.

The catalog

Section titled “The catalog”

embed::catalog() returns &'static [fetch::FetchSpec], one entry per file above, all from onnx-community/embeddinggemma-300m-ONNX at commit 5090578d9565bb06545b4552f76e6bc2c93e4a66:

Field Type Holds
url &'static str the file’s download address, pinned to the commit
dest_rel &'static str models/embeddinggemma-300m-q4/ followed by the file’s path in the folder above
sha256 &'static str the file’s sha256, 64 lowercase hex characters
size_hint Option<u64> the file’s size in bytes, always present
File Bytes sha256
config.json 1,765 6e1f06404b7163e0325ed2ea3e6781cde50f4a50b31780a95ad0d30e8404d77b
special_tokens_map.json 662 2f7b0adf4fb469770bb1490e3e35df87b1dc578246c5e7e6fc76ecf33213a397
tokenizer_config.json 1,156,830 3ca953eea6c3c9fcda9cf3df22949ff18b216f7c74bd6459230f3f1013953f3a
tokenizer.json 20,323,312 4dda02faaf32bc91031dc8c88457ac272b00c1016cc679757d1c441b248b9c47
onnx/model_q4.onnx 519,322 ad1dfee81a70f7944b9b9d1cc6e48075b832881cf33fab2f2b248be78f3f0043
onnx/model_q4.onnx_data 196,725,760 599962c3143b040de2dd05e5975be3e9091dd067cacc6a8f7186e3203bab9e02

The examples’ model folder

Section titled “The examples’ model folder”

The search example keeps the model in $PITO_EMBED_MODELS when it is set, otherwise in $XDG_DATA_HOME/pito-embed/embeddinggemma-300m-q4 (~/.local/share/pito-embed/embeddinggemma-300m-q4 when that isn’t set) on Linux and ~/Library/Application Support/pito-embed/embeddinggemma-300m-q4 on macOS. The bench example and the tests that run the real model read PITO_EMBED_MODELS.