ShapedQL

A second key capability lies in the rich set of embedding types available for retrieval. In the ShapedQL Playground, the demo model includes title_embedding for semantic matching of movie titles, description_content_embedding for plot-based relevance, and poster_embedding for image similarity searches. Additionally, collaborative_embedding and people_also_liked embeddings enable user-item interaction-based recommendations, while personnel_embedding supports diversity and filtering by cast and crew. Each embedding can be referenced directly in a ShapedQL query, allowing developers to mix and match heterogeneous signals—semantic text, visual content, and behavioral patterns—within a single retrieval expression. This modularity means that a query can, for instance, retrieve items visually similar via poster_embedding and then boost those with high collaborative affinity, all without writing separate pipelines.

Beyond retrieval, ShapedQL includes advanced scoring and hybrid search capabilities. The playground features a click_through_rate model trained using LightGBM, which can be invoked in queries to order results by predicted user engagement. This learned scoring can be combined with a personalized reranker that adapts to individual user profiles. The language also natively supports hybrid search, blending lexical keyword matching with semantic vector similarity to improve recall and precision. By integrating these ranking and personalization features directly into the query, ShapedQL eliminates the need for separate post-processing steps, making the final output more directly aligned with business metrics like clicks or conversions. Moreover, the scoring models can be swapped or weighted dynamically, allowing fine-grained control over the trade-off between relevance and personalization. This integration ensures that the entire pipeline from retrieval to final ranking is expressed in a single, maintainable unit.

The ShapedQL Playground provides a hands-on environment to experience this workflow. It comes preloaded with the Movielens dataset, enriched with IMDb metadata, so users can immediately start experimenting. The interface guides you through selecting a use case—Agent retrieval, Search and feeds, or Recommendations—and then writing or selecting a ShapedQL expression. Upon execution, the query engine processes the expression against the live embeddings and scoring models, returning ranked results in the right-hand pane. This immediate feedback loop allows practitioners to gain intuition about embedding behaviors and scoring nuances, accelerating the development of production-grade retrieval pipelines. Because the dataset and embeddings are preconfigured, there is no setup overhead; users can focus entirely on query logic and ranking strategies. The playground also supports saving and reusing queries, making it easy to collaborate and share best practices across a team.

Concrete real-world scenarios illustrate the language’s versatility. A developer building a movie streaming service can use description_content_embedding to find films with plots similar to a user’s recently watched titles, then apply collaborative_embedding to boost items popular among similar users, and finally rank by click_through_rate to maximize engagement. An e-commerce team can employ hybrid search with lexical and semantic matching to handle a wide range of user queries, while a chatbot creator can leverage agent retrieval to fetch contextually relevant movie information based on natural language. The result in each case is a more tailored and effective search or recommendation experience, built with less code and fewer moving parts. Furthermore, a media company could utilize poster_embedding for visual browsing of movie thumbnails, allowing users to discover content by image similarity, while the personnel_embedding helps surface films featuring favorite actors or directors. Each scenario benefits from ShapedQL's ability to unify retrieval, filtering, and scoring in a single, auditable query.

ShapedQL is intended for ML engineers, data scientists, and software developers who are tasked with delivering high-quality search and recommendation features. Whether working on a consumer app, an internal tool, or a large-scale platform, these professionals benefit from the language’s unification of retrieval and ranking. The playground is entirely web-based, requiring no local setup, and the technology is designed for seamless integration into existing tech stacks via APIs. In summary, ShapedQL revolutionizes recommendation system development by combining the clarity of SQL with the power of modern embeddings and machine-learned scorers, enabling teams to create more intelligent and engaging user experiences faster. By empowering these roles to express complex ideas in a concise, readable format, ShapedQL fosters cross-functional collaboration and accelerates the journey from data to delight. Its immediate applicability to common search and feeds use cases makes it an essential tool for any organization serious about personalization.