Same hybrid pipeline as baseline-v1-desc-hybrid, but the encoder is sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2 (117M params, 50+ languages). ~80 MB of centroids on disk. Built for agents that receive tasks in French or any of the 50+ languages supported by the encoder.

En français : ce modèle route une tâche écrite en français vers les bons outils parmi un catalogue de 18 671 outils anglophones (les descriptions des outils restent en anglais, l'encoder multilingue aligne les deux espaces).

Quick start

Install the SDK directly from GitHub (PyPI publish pending):

pip install git+https://github.com/dalek-ai/agent-tool-router.git

pip install "agent-tool-router[encoder] @ git+https://github.com/dalek-ai/agent-tool-router.git"

python

from agent_tool_router import Router
r = Router.from_pretrained("baseline-v1-desc-hybrid-multilingual")
r.route("envoie un message slack à l'équipe data", k=3)
# ['post_message_to_slack', 'sendSlackMessage', ...]
r.route("traduis ce document en anglais", k=3)

Numbers

Parallel EN/FR probe (n=50 hand-written queries against the full 18 671-tool catalog, top-3 per call):

model	EN top-3	FR top-3	EN top-5	FR top-5
baseline-v1-desc-hybrid (default English encoder)	82%	26%	90%	30%
baseline-v1-desc-hybrid-multilingual	82%	54%	90%	62%

+28pp top-3 on French queries with no measurable cost on English top-3 at this panel size. Reproduce: python -m router.eval.eval_fr_pretrained.

English-only benchmark (LOSO refit, full catalog, top-3): the multilingual encoder trails the default English-only encoder by ~3.9pp weighted overall (Hermes -8.8pp, ToolACE -3.9pp, tau-bench -1.5pp). Prefer the default model if all your queries are English.

Repo & demo

github.com/dalek-ai/agent-tool-router · MIT. Live demo: dalek-ai/agent-tool-router-demo (gradio Space, FR/EN).

dalek-ai/baseline-v1-desc-hybrid-multilingual

baseline-v1-desc-hybrid-multilingual

Quick start

Numbers

Repo & demo

dalek-ai/
baseline-v1-desc-hybrid-multilingual