Show HN: Shaped – Fine-tuning semantic search on behavioral signal

When OpenAI released Learning Transferable Visual Models From Natural Language Supervision in 2021, it sparked a multimodal semantic search revolution. While the concept wasn’t new, the paper catalyzed a shift in the industry, inspiring a wave of semantic search tools. These advances delivered a leap forward in search quality compared to traditional lexical search, with adopters reporting significant improvements in conversion rates and revenue.

But semantic search alone isn’t the endgame. Productionizing search remains a challenge, and hybrid search—combining keyword precision with semantic depth—is still found in the state-of-the-art for relevancy. The next frontier lies in sophisticated reranking techniques, particularly learning-to-rank (LTR) models. By leveraging behavioral signals and user preferences, these approaches optimize search outcomes far beyond just query relevance, enabling personalized and adaptive experiences. Such models dynamically tailor search results based on historical interactions, user context, and business objectives, balancing relevance, diversity, and fairness. The result? Smarter, more impactful search that drives better user experiences and business outcomes.

We built Shaped on the belief that LTR is the future of content & product discovery, and that includes LTR for search. Taking the benefits of personalization as an example – why should every user get the same results when their needs and behaviors differ?

Shaped is a platform designed to bring AI-native search to the 99%. With an emphasis on easy integration, rapid experimentation, and flexible configurability, developers can deploy advanced search and recommendation systems in less than a sprint.

Here's how it works:

Connect your datastack, including multi-modal item catalog and optional real-time event streams. Choose your objective (e.g., clicks, new-user activation) using a flexible value-model interface. By default, Shaped selects the optimal search architecture, including search type and reranking models, however, you can configure all of this manually. Shaped deploys and orchestrates the infrastructure, models and pipelines needed for a scalable, AI-native search engine with continuous indexing and learning-to-rank model training. Access a UI for performance monitoring and result exploration.

Check out our demo sandbox at (play.shaped.ai) and try the text query input. We’d love your feedback and thoughts on how to push AI powered search further!