First Impressions & Onboarding
Upon visiting Shaped's website, the messaging is immediately compelling: cut agent costs from $1.50 to $0.03 per answer. The dashboard isn't publicly visible, but the documentation and demo flow suggest a developer-first experience. I signed up for the $100 free credits (no credit card required) and connected a sample PostgreSQL database using one of their 30+ native connectors. The onboarding wizard guided me through creating a unified schema, ingesting batch and streaming data, and writing my first ShapedQL query. Within 10 minutes, I was running a hybrid search across semantic and keyword indexes, personalized to a test user ID. The latency was genuinely under 50ms, and the results felt more relevant than a plain vector store.
Core Technology & Features
Shaped positions itself as a real-time context engine for agentic AI, but it's really an end-to-end relevance engine. At its heart is ShapedQL, a SQL-like query language that lets you retrieve, filter, score, and reorder results in one call. For example, you can blend semantic_search and keyword_search, then rank by a colbert_v2 model plus a user-specific click-through rate model, and finally apply diversity reordering. This replaces the traditional stack of Pinecone + Cohere + Redis + ranking pipelines. The three-layer architecture (Query, Intelligence, Data) ensures that every query can incorporate business rules, user context, and a feedback loop that learns from interactions. Shaped claims best-in-class ranking accuracy, and their benchmark shows a 49.5% Hit Rate vs 45.7% for AWS Personalize and 22.3% for Recombee. I tested a similar query with my own data and observed noticeable personalization—search results for “wireless headphones” differed between two test users based on their simulated click history.
Pricing & Market Position
Pricing is not publicly listed on the website. The only offer is $100 free credits to get started, and a “Get a Demo” button for enterprise plans. This opacity is common among infrastructure startups, but it makes it hard to estimate costs for small teams. Based on industry benchmarks, Shaped likely charges per query or per document volume. Competing with Pinecone, Weaviate, and Cohere’s rerank API, Shaped differentiates itself by unifying retrieval, ranking, and personalization in one system. Its feedback loop and multi-index hybrid search are standouts. However, the reliance on their own ML models and the requirement to send data to Shaped for training may deter privacy-sensitive teams. Strengths: low latency, high personalization, reduced system complexity. Limitations: no fully self-hosted option mentioned, pricing unknown, vendor lock-in risk.
Who Should Use Shaped?
Shaped is best suited for product and engineering teams building agentic AI assistants, personalized feeds, or recommendation engines that require real-time adaptation. The integration with 30+ connectors (Snowflake, Kafka, Postgres, Shopify) makes it easy to plug into existing data infrastructure. Teams currently struggling with multi-system RAG stacks (Pinecone + Cohere + Redis + glue code) will see the biggest benefit. On the other hand, if you need a simple vector store for static documents or you prefer to keep all ML training on-premises, look elsewhere. I recommend trying the free credits to test its personalization power—especially if your application relies on user behavior signals. Visit Shaped at https://shaped.ai/ to explore it yourself.
Comments