baseline-v1-desc

Centroid retrieval router over 18 671 tools, scored by cosine(task_tfidf, description_tfidf). TF-IDF only, ~6 MB, no torch, no GPU. The default model behind pip install agent-tool-router.

Quick start

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

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

from agent_tool_router import Router
r = Router.from_pretrained("baseline-v1-desc")
r.route("cancel my pending order and refund the credit", k=3)
# ['refundOrder', 'modify_pending_order_items', 'cancel_pending_order']

Numbers

Per-call top-3 against the full 18 671-tool catalog (n=30 425 calls, held-out across the corpus). Random baseline = 3/V = 0.016%:

source	n calls	top-3
Hermes function-calling-v1	4 376	74.3%
ToolACE	17 169	52.4%
tau-bench	8 880	3.2%
overall	30 425	41.2%

For Pareto-better top-3 across all three sources at the cost of ~250 MB of torch + sentence-transformers, switch to baseline-v1-desc-hybrid.

Repo & demo

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