How I Built & Shipped Tool Recommendation
Blog post from Arcade
Arcade.dev’s Tool Recommendation system was developed over nine months to help AI agents efficiently access large tool catalogs without processing every available option, reducing token use and improving performance. Rather than relying on expensive generative models to identify relevant tools, it uses specialized embedding and ranking models to retrieve the most suitable tools before passing them to the primary reasoning model, potentially allowing less expensive models to perform more effectively for tool-based tasks. Although the core recommendation model performed well early on, production deployment required addressing enterprise concerns including policy integration, on-premises and air-gapped environments, GPU availability, hardware configurations, privacy assurances, and deployment architecture. The team also found that seemingly minor details, such as exposing tool version numbers or changing formatting used in re-ranking, could cause schema failures or lower retrieval accuracy; the system now reportedly achieves nearly 98% Recall@10. Arcade argues that its approach differs from simpler text-search implementations because it is designed for enterprise-scale reuse and cost efficiency, and it plans to extend the retrieval framework beyond tools to organizational knowledge and context.
| Trend | Post Mentions | Total Month Mentions | Posts | Companies | MoM |
|---|---|---|---|---|---|
| Vector Search | 2 | 1,978 | 415 | 135 | +18% |
| AI Model Fine-tuning | 1 | 868 | 227 | 77 | +17% |
| LLM | 1 | 6,578 | 1,243 | 236 | +9% |
Use this post, company, and trend context to find content marketing opportunities, perform competitive analysis, or address product feature gaps via the Plushcap MCP server or the Plushcap API.