Module models 

Where the weights come from.

A published crate cannot carry its models. crates.io caps a package at 10 MB and the full model set is ~388 MB, so the crate ships what genuinely fits and resolves the rest at runtime.

The split is not arbitrary, and one half of it is a licensing constraint rather than a size one:

what	size	how it arrives
trained heads and the reasoning core	~1.5 MB	compiled into the crate
relation head	~2 MB	compiled into the crate
the bert-tiny embedding table	15.2 MB	downloaded — too large to package
our tagger / SPLADE / LoRA weights	100s of MB	downloaded from our own model repo
third-party models (needle3, model2vec, PP-OCR)	100s of MB	fetched from their repos, never re-hosted

That last row is the important one. We trained the tagger, the facet heads and the LoRA, so we may redistribute them. needle3 belongs to Cactus, bert-tiny to Google, and the others to their authors — re-hosting someone else’s weights under our name would be wrong regardless of whether it is convenient, so those are fetched from the repository that published them and the licence is recorded here.

Nothing downloads without being asked: resolve searches locally first and returns ModelError::NotFound with the exact fetch instruction if it comes up empty.

Modules

bundled

The trained parameters small enough to travel with the crate.

Structs

Artifact

One resolvable set of weights.

Enums

ModelError

Why a set of weights could not be produced.

Source

Where a set of weights comes from, and what we are permitted to do with it.

Constants

ARTIFACTS

Every artifact the engine knows how to find.

Functions

artifact

Look up an artifact by name.

fetch_hint

The instruction that would make a missing artifact present.

resolve

Find the directory holding an artifact, without downloading anything.

search_roots

Directories searched for models, in order.

status

A short report of what is present and what is missing — useful in a CLI or a bug report.