HopitAI/moda-duo

MODA Duo

Two open constituents, one answer per query. Duo routes each text query to whichever open MODA system suits its shape — short catalogue titles to MODA Pro Lite+, longer descriptions to MODA — and runs one encoder and one nearest-neighbour query per search.

Duo adds zero parameters. It is a serving recipe over two open systems, not a new model.

Why

Fashion search queries come in two shapes, and no single small model is best at both:

query shape	example	best open system ≤250M
catalogue title	buckle round toe flat shoes	MODA Pro Lite+
natural description	When warm weekends are abound, make sure your closet…	MODA

Duo picks per query. On a mixed workload it beats both constituents.

Results

MAP@10, full corpus, all ground-truth queries, one evaluator (pytrec_eval map_cut.10), paired bootstrap 10,000 resamples.

benchmark	MODA	MODA Pro Lite+	MODA Duo
KAGL	0.2887	0.3201	0.3201
Polyvore	0.3726	0.4049	0.4049
Atlas	0.1862	0.1904	0.1904
Fashion200K	0.1946	0.1846	0.1866
DeepFashion In-Shop	0.1642	0.1026	0.1640
DeepFashion Multimodal	0.0147	0.0133	0.0159
pooled, 12,000 queries	0.2035	0.2026	0.2137

Pooled across all six benchmarks — the mixed workload a router exists for — Duo is +5.0% over MODA and +5.4% over MODA Pro Lite+, both significant.

Fashion200K is the honest miss: its queries sit where the two constituents are hardest to tell apart, and Duo trails MODA there by 4%. Where a workload is known to be all long descriptions, use MODA directly.

Serving cost

indexes                  2     one per constituent, built offline
stored vectors per item  2
encoders run per query   1     only the routed constituent's text tower
ANN queries per search   1
re-ranking               none

Compared with a single open model, Duo costs one extra index at build time and nothing extra at query time.

Use

pip install open_clip_torch pillow numpy hnswlib
python serving_ann.py --demo

from serving_ann import Duo
duo = Duo()                 # loads both constituents
duo.build(images)           # encodes the catalogue with both, builds two indexes
ids, scores, routes = duo.search(["black leather ankle boots"], k=10)

The router is a callable — replace it with any policy that maps a query to a constituent:

duo = Duo(router=lambda q: "moda" if looks_like_a_description(q) else "moda_pro_lite_plus")

Evaluation

All figures are full corpus, all ground-truth queries, MAP@10 under one evaluator (pytrec_eval map_cut.10), paired bootstrap with 10,000 resamples. Per-query results are in the repository.

Related

• MODA — FashionSigLIP with a serving harness. Open source, open weights.
• MODA Pro Lite — a trained fashion encoder. Open weights.
• MODA Pro — hosted. Fuses both constituents rather than choosing between them.