# Choose How to Run Relevance-Based Prediction
![[Pasted image 20260811010135.png]]
Cambridge Sports Analytics offers Relevance-Based Prediction through two production-grade deployment models: the local RBP Math Library and the hosted Prediction Engine API.
Both use the same Relevance-Based Prediction method and provide transparent predictions with relevance, fit, asymmetry, weights, and other supporting insights. The best choice depends on how you want to deploy, scale, and operate RBP.
## RBP Math Library
The RBP Math Library is a compiled native engine that runs directly inside your Python or native application.
Choose the local library when:
* Data must remain inside your workstation, server, VPC, or on-premise environment.
* Performance and low per-call latency are priorities.
* Compliance or security requirements call for direct control over execution and data flow.
* You need embedded, private, or offline operation.
* Your team wants to choose the hardware and manage application-level concurrency.
Local execution avoids network requests, serialization, queues, and service round trips. It can provide substantially higher performance on suitable hardware and workloads.
[Explore the RBP Math Library](/preview/site_vZwTs/~/revisions/nzLvNosuNDm5V5Bpvu9B/rbp-math-library/overview.md)
## Prediction Engine API
The Prediction Engine API provides RBP as a managed HTTP service with a Python client.
Choose the API when:
* You prefer API-based integration and managed service operations.
* You need broad, elastic concurrency beyond the hardware available to a local process or server.
* Many applications, users, or teams need a common prediction endpoint.
* Asynchronous submission and centralized result retrieval fit your architecture.
* You want to use RBP without deploying native binaries in each application environment.
[Explore the Prediction Engine API](/preview/site_vZwTs/~/revisions/nzLvNosuNDm5V5Bpvu9B/prediction-engine/overview.md)
## Comparison
<table data-header-hidden="false" data-header-sticky data-search="false"><thead><tr><th>Consideration</th><th>RBP Math Library</th><th>Prediction Engine API</th></tr></thead><tbody><tr><td>Execution</td><td>Inside your application or infrastructure</td><td>Hosted service</td></tr><tr><td>Data flow</td><td>Remains in your environment</td><td>Sent securely to the API</td></tr><tr><td>Performance</td><td>Lowest latency and native local performance</td><td>Includes network and service round trips</td></tr><tr><td>Concurrency</td><td>Determined by your hardware and architecture</td><td>Managed, broad elastic concurrency</td></tr><tr><td>Operations</td><td>Managed by your team</td><td>Managed by CSA</td></tr><tr><td>Integration</td><td>Python package or C interface</td><td>REST API or Python client</td></tr><tr><td>Offline use</td><td>Supported</td><td>Requires service connectivity</td></tr><tr><td>Enterprise use</td><td>Compliance, security, performance, embedded and on-premise workloads</td><td>Shared services, managed scaling, centralized access and high concurrency</td></tr></tbody></table>
> [!NOTE]
> These options are complementary. An organization can use the local library for latency-sensitive or regulated workloads and the API for broadly concurrent or shared workflows.
## Learn the Method
If you are new to the underlying approach, continue to [Relevance-Based Prediction](/preview/site_vZwTs/~/revisions/nzLvNosuNDm5V5Bpvu9B/relevance-based-prediction/glossary.md) for concepts, terminology, and further reading.