Multilingual base — encoder is sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2 (117M params, 50+ languages), so French / Spanish / German queries route into the same 18 671-tool English catalog as the default.
Next-tool fine-tune — the multilingual encoder has been fine-tuned on 8 386 next-tool prediction triplets with a contrastive (task, gold_description, hard_negative) objective. Hard negatives come from the previous hybrid encoder's top-50 retrieval, so each gradient step pushes past its own retrieval mistakes.

Encoder lives at dalek-ai/multilingual-next-v1, loaded lazily on the first route() call. ~33 MB of TF-IDF + encoder centroids on disk in this repo.

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-next-v1")
r.route("annule ma commande en attente et rembourse le crédit", k=3)
r.route(
    "I want to add a checked bag to my reservation",
    k=3,
    history=["update_reservation_flights"],
)

Numbers

Next-tool prediction on 2 094 held-out triplets (trace_id 80/20 split, seed=17, Markov-1 train-only):

Encoder	recall@50	recall@200	Markov top-3 K=50	Markov top-3 K=200
`MiniLM-L6` (baseline-v1-desc-hybrid)	58.9%	69.6%	48.0%	54.9%
`multilingual-MiniLM-L12` (baseline-v1-desc-hybrid-multilingual)	—	—	—	—
`minilm-next-v1` (baseline-v1-desc-hybrid-next-v1)	83.4%	93.1%	70.2%	75.5%
`multilingual-next-v1` (this model)	85.2%	92.1%	71.5%	73.7%

The L12 base beats the L6 base on recall@50 and Markov top-3 K=50, at the cost of ~2pp on recall@200 (the L6 fine-tune still wins the top-200 rerank). Reproduce:

python -m router.eval.eval_next_tool_widen --cache-dir data/cache/next_tool_multilingual_v1 --model dalek-ai/baseline-v1-desc-hybrid-multilingual-next-v1

LOSO refit hybrid (full 18 671-tool catalog, α=0.5, top-3 per call):

held_out	n_calls	`MiniLM-L6`	multilingual base	`minilm-next-v1`	multilingual-next-v1
Hermes function-calling-v1	4 376	72.3%	63.5%	73.6%	70.8%
ToolACE	17 169	58.7%	54.8%	64.8%	59.1%
tau-bench	8 880	11.1%	9.6%	38.8%	43.5%

The fine-tune Pareto-dominates the plain multilingual base on every held-out source (Hermes +7.3pp, ToolACE +4.3pp, tau-bench +33.9pp). Compared to the English-only minilm-next-v1 fine-tune, it gives up ~3pp on Hermes / ToolACE in exchange for an extra +4.7pp on tau-bench and the multilingual coverage.

History bigram rerank (Markov-2, ≥ 0.4.0)

Passing a 2-step history triggers stupid backoff to the bigram table shipped in this repo (markov2_counts.npz + markov2_keys.npy, ~25 KB), falling back to Markov-1 when the (prev2, prev1) pair is unseen in train. Same held-out 2 094 triplets, K=200, best alpha per system:

system	top-1	top-3	top-5
Markov-1 only (≥ 0.3.0)	51.7%	73.7%	80.1%
Markov-2 backoff (≥ 0.4.0)	56.9%	76.2%	80.5%
delta	+5.2pp	+2.5pp	+0.4pp

Largest single-position gain: tau-bench t=2 reaches 100% top-3 with Markov-2 (was 97.4% with Markov-1). tau-bench t≥3 climbs to 92.9% top-3 (+8.1pp). On the 775 test rows whose (prev2, prev1) bigram is seen in train, top-1 jumps +14.0pp (62.5% → 76.5%). Reproduce:

python -m router.eval.eval_next_tool_markov2 --cache-dir data/cache/next_tool_multilingual_v1 --model dalek-ai/baseline-v1-desc-hybrid-multilingual-next-v1

FR/EN qualitative probe (n=50 parallel queries)

model	EN top-3	FR top-3	EN top-5	FR top-5
baseline-v1-desc-hybrid (default `MiniLM-L6`)	82%	26%	90%	30%
baseline-v1-desc-hybrid-multilingual	82%	54%	90%	62%
baseline-v1-desc-hybrid-next-v1	86%	28%	94%	28%
baseline-v1-desc-hybrid-multilingual-next-v1	84%	54%	90%	60%

No French drift from the EN-only fine-tune triples (54% top-3 FR identical to the plain multilingual). +2pp EN top-3 over the plain multilingual. Reproduce: python -m router.eval.eval_fr_pretrained.

When to use which

Catalog / queries	Recommended model
English-only queries	`baseline-v1-desc-hybrid-next-v1`
Mixed FR/EN, multi-step agents	this model
FR/EN without history rerank	`baseline-v1-desc-hybrid-multilingual` (strictly dominated by this model, kept for backward compat)

Repo & demo

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