Shaped review

Shaped is a fully managed relevance service: a hosted platform that sits between an application's data sources and the code that needs ranked results. It is not presented as a library or a self-hosted server. The service connects data, creates embeddings, retrieves candidates, ranks them with user and item context, and can use interactions as feedback for future ranking.

That places Shaped across parts of a typical search or recommendation stack that teams might otherwise implement with separate indexing, retrieval and ranking components. Its use cases include hybrid search, personalized feeds, product recommendations and retrieval for agents or RAG applications. ShapedQL is the query interface for retrieval, filtering, scoring and reordering. The homepage says results can be returned in under 50ms; treat that as the vendor's claim, not a published service-level guarantee. Data can arrive through batch and streaming connectors, then be exposed through unified schemas and SQL transforms.

The main query surface is ShapedQL, a SQL interface that combines retrieval functions with filters, ranking scores and reorder operations. The homepage shows semantic and keyword search in one query, with examples using a reranking model, click-through-rate model and diversity function. It also says results can be retrieved by text, user ID or item ID.

The homepage lists a Python SDK, a TypeScript SDK and MCP. It does not name additional programming languages or a CLI. The integration surface also includes 30+ native connectors for batch and streaming data; named sources include Snowflake, BigQuery, Redshift, Databricks, Amplitude, Segment, RudderStack, PostHog, Kinesis, Kafka, Pub/Sub, Postgres, MongoDB and Shopify. Check Shaped's docs for request transport, authentication, SDK coverage and connector-specific setup before implementation. The homepage links to docs but does not describe deployment workflows or a local development environment.

Starter is free and lists console access, community support, no credit card requirement and free credits. The pricing material lists those credits as both $100 and $300 per month, so confirm the actual allowance and any expiry terms. It also lists a 14-day free trial. Annual pricing is not published.

Standard costs $42 per million queries and has a $500 monthly minimum spend. GPU training and encoding are listed at $5 per hour, while storage is $0.75–$2.25 per GB. The pricing material also includes earlier rates of $23 per million calls, $3 per GPU hour and $0.75–$1.50 per GB. Ask which rates apply to a new account and how usage is measured. Pro support is listed at $500 per month. Enterprise pricing is custom, with dedicated engineering, custom SLAs and 24/7 support available. For production budgeting, model query volume, GPU time and stored data separately; the Standard minimum applies even when usage charges are lower.

Shaped is described as a fully managed cloud platform. The vendor does not publish a software license or source-availability terms in the published product details, so do not assume the service is open source. Self-hosting options are also not published; teams that require deployment in their own environment should ask Shaped directly.

The vendor does not publish data export formats, export procedures or migration tooling in the published details. That matters because a Shaped integration can involve its unified schemas, embeddings, ranking models, behavioral signals and ShapedQL queries. Before adoption, establish which source data and derived artifacts can be exported, in what formats, and whether models or ranking logic can be reproduced elsewhere. Keeping canonical event and catalog data in systems you control will reduce migration risk, but does not by itself preserve Shaped-specific ranking behavior.

The homepage claims results in under 50ms and lists SOC 2 Type II compliance. It also describes continuous learning and batch plus streaming connectors. Those points make the platform relevant to latency-sensitive ranking, but the vendor does not publish API rate limits, request-size limits, data-retention terms or detailed latency conditions in the published material. Ask how the latency claim is measured and what happens during overload or connector lag.

The pricing information lists a 99.95% uptime SLA as available for Enterprise, not as a default commitment for every plan. Standard has a monthly minimum and usage-based query, compute and storage charges; confirm alerts, quotas and overage handling before launch. Teams should also verify incident response, backup and recovery terms, data deletion procedures and regional availability. The stated SOC 2 Type II status is a useful security signal, but it does not answer those operational questions or replace a review of the vendor's security documentation.