Module embed 

Which embedding model a store uses, and how to load it.

fastembed ships an enum of models it knows how to fetch and configure. The default is not in that enum: granite-embedding-small-english-r2, quantized to int8, has to be assembled by hand from its ONNX graph, its separate weights file, and its tokenizer. Model is the union of “something fastembed knows” and “the one we assemble ourselves”, so a store can record either and read it back.

Enums

Model

A model a store can be built with.

Constants

AIRGAP_ENV

How many threads ONNX Runtime should use inside one operator.

GRANITE_FILES

Every file semlith fetches from [GRANITE_REPO], with its SHA-256.

GRANITE_NAME

What a store writes into its model meta row for granite. Changing this string orphans every store that recorded the old one.

GRANITE_REVISION

The commit each model repository is fetched at.

THREADS_ENV

Cap on ONNX Runtime’s intra-op thread pool, overriding the derived value.

Functions

airgap

check_cache_dir

Refuse a model cache that somebody else on this machine can write to.

digest

The SHA-256 of some bytes, as lowercase hex.

embed_threads

is_cached

Whether the default model is cached, with the right bytes.

link_runtime

Point ONNX Runtime at the library shipped beside this binary.

model_is_cached

Whether one model’s weights are already in the cache.

refuse_if_airgapped

Whether the cache already holds something to load.

verify

Check bytes against the digest recorded for them, or say which file failed.

verify_cached

Verify every pinned file of a repository that fastembed fetched itself.