Encoder = sentence-transformers/all-MiniLM-L6-v2 (22M params, English). Encoder is lazy-loaded on the first route() call. ~35 MB of centroids on disk.

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")
r.route("cancel my order and refund the credit", k=3)
# ['cancel_pending_order', 'cancel_order', 'refundOrder']

History-aware rerank (≥ 0.3.0)

The model ships a Markov-1 transition table (~21 KB) trained on the same 7 184 multi-turn traces. Pass the tool names already called in the trace as history= and the top-200 retrieval candidates are reranked with a learned prior:

python

r.route(
    "I want to add a checked bag to my reservation",
    k=3,
    history=["update_reservation_flights"],
)
# ['update_reservation_baggages', 'update_reservation_passengers', 'cancel_reservation']

Measured lift on n=2094 held-out triplets (split by trace_id, train-only Markov to avoid prior-on-test leakage; alpha sweep-best per bucket):

Setup	top-1	top-3	top-5
Retrieval-only	13.8%	32.7%	38.8%
Markov-1 rerank top-50 (α=0.4)	34.6%	48.0%	50.5%
Markov-1 rerank top-200 (α=0.1) ⬅ default ≥ 0.3.0	39.0%	54.9%	57.7%

Widening the retrieval bucket from 50 → 200 lifts top-3 by +6.7pp without any new training: same prior, more candidates available to rerank. Retrieval recall@K is the mechanical ceiling: recall@50 = 58.9%, recall@200 = 69.6% on this test split.

No-op when history is omitted or empty. Override with markov_alpha=0.0 (prior-only) or 1.0 (retrieval-only), and the bucket width via markov_rerank_n (default 200).

Numbers

Per-call top-3 against the full 18 671-tool catalog (n=30 425 calls):

source	n calls	tfidf	encoder	hybrid α=0.5
Hermes function-calling-v1	4 376	74.3%	60.7%	74.9%
ToolACE	17 169	52.4%	54.8%	62.8%
tau-bench	8 880	3.2%	6.1%	9.9%
overall	30 425	41.2%	41.4%	49.1%

The hybrid Pareto-dominates both solo backends on every source and every top-k (+7.9pp overall vs TF-IDF). LOSO refit (TF-IDF retrained per held-out source, encoder pretrained) loses at most -4.1pp vs the in-distribution number above.

For routing on French / non-English tasks, see baseline-v1-desc-hybrid-multilingual.

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

baseline-v1-desc-hybrid

Quick start

History-aware rerank (≥ 0.3.0)

Numbers

Repo & demo

dalek-ai/
baseline-v1-desc-hybrid