madosho-cn-oth-src

A permissively-licensed bundle of non-allied RAG components for madosho. Everything here is origin_tier = CN_ORIGIN (BAAI, Alibaba) but permissively licensed (MIT / Apache-2.0). The code forces no license on madosho; the reason this is a separate package is operator opt-out: an operator whose policy excludes non-allied components simply does not install this bundle, and the whole batch disappears from the component menu at once. madosho's in-tree baseline stays allied.

This is the permissive + non-allied quadrant of madosho's component taxonomy (the other off-baseline quadrants are the AGPL allied-agpl parser bundle and a future copyleft + non-allied bundle). origin_tier is a label madosho records honestly; it never gates selection.

Components

kind	name	org	country	license
embedder	bge-base-en-v1.5	BAAI	CN	MIT
embedder	qwen3-embedding-0.6b	Alibaba	CN	Apache-2.0
reranker	bge-reranker-v2-m3	BAAI	CN	Apache-2.0
reranker	bge-reranker-base	BAAI	CN	MIT

Each is a thin subclass of madosho's public StEmbedder / StCrossEncoderReranker base classes; only the default model id + metadata card differ. Model weights download on first use into the shared HF cache.

Install (host venv, for tests/CLI)

pip install --no-deps -e .

(madosho and sentence-transformers are expected already present, as in the madosho dev venv.)

Install (compose overlay)

Chain after the AGPL bundle so one dev image lists everything:

docker build --build-arg BASE=madosho-allied-agpl:local -t madosho-plugins:local .

or stand alone on the clean base with --build-arg BASE=madosho:local.

Verify it loaded

Once installed, madosho discovers all four through the madosho.components entry point; they show up in madosho components list with the origin and license madosho records for them (labels, never gates):

kind      name                  license     org      origin      hardware  extra
embedder  bge-base-en-v1.5      MIT         BAAI     cn_oth_src  cpu       -
embedder  qwen3-embedding-0.6b  Apache-2.0  Alibaba  cn_oth_src  cpu       -
reranker  bge-reranker-base     MIT         BAAI     cn_oth_src  cpu       -
reranker  bge-reranker-v2-m3    Apache-2.0  BAAI     cn_oth_src  cpu       -

Test

pytest -q

Tests prove registry discovery + class loading; they do not download weights.