Researchers adapt embedding models to improve agent capability retrieval in marketplaces
Fine-tuned BGE-base, KaLM, EasyRec models boost search across 1,000 queries in unseen agent catalogs
Open agent marketplaces are growing rapidly, hosting a mix of native agents, tool bundles, and reusable skill packages—all accessible through a single search interface. However, practitioners face a key challenge: how to retrieve relevant executable capabilities when queries are diverse and catalog items vary widely in format and functionality. Current off-the-shelf retrieval models, designed for generic text matching, lack the specialized understanding needed to match user intents to agent capabilities.
A team led by Tingwei Chen proposes a solution: fine-tuning existing embedding models on a curated dataset called AgentSelect, which converts marketplace-visible units into structured capability profiles. They fine-tune three open backbones—BGE-base, KaLM-v1.5, and EasyRec—and test transferability on two previously unseen catalogs: MuleRun native agents and ClawHub (50 skills with 1,000 queries). Results show that even lightweight adaptation significantly improves retrieval accuracy across both catalogs. Accepted for oral presentation at the AgentSearch Workshop at SIGIR 2026, the work provides a practical path for building better search in agent ecosystems, with code and data promised upon publication.
- Fine-tuned three open embedding models (BGE-base, KaLM-v1.5, EasyRec) on AgentSelect, a dataset of marketplace capability profiles.
- Transfer tested on two unseen catalogs: MuleRun native agents and ClawHub (50 skills, 1,000 queries).
- Accepted at AgentSearch Workshop, SIGIR 2026; code and data to be released.
Why It Matters
Enables practitioners to effectively search across mixed agent catalogs, accelerating discovery and reuse of capabilities.