Build a platform on Splendor

If you are building a product on top of Splendor — a screenshot search app, a document search feature, an analytics layer for your own customers — you are a platform. Your customers each have their own data that must never be visible to another customer. This guide shows how to model that.

​

The one rule: a customer is a tenant, not a dataset

A tenant is Splendor’s hard isolation boundary. Every query, handle, and stored record is scoped to one tenant, and the API never lets data cross that line — it is enforced at the auth, query, handle, and storage layers, independent of your application code. A dataset is a namespace within a tenant. Datasets are co-queryable: a single search can target several at once, and there is no per-dataset access control. They organize kinds of data, not owners of data. So the mapping for a platform is:

​

How access works

Your backend holds a platform API key (a machine-to-machine credential). With it you can:

• Provision a tenant per end-customer (POST /v1/platform/tenants).
• Act inside any of your tenants by sending its id in the X-Splendor-Tenant-Id header alongside the platform key — search, ingest, export, everything.
• Tear a tenant down completely when a customer leaves (DELETE /v1/platform/tenants/{id}).

A platform key only ever reaches the tenants you created. It cannot reach another platform’s tenants, and it is not a Splendor staff credential. Your end-users never authenticate to Splendor directly — your application is the only client. See Provision tenants for your customers for the end-to-end flow.

​

Worked example: a screenshot search app

Say you run a screenshot search product. Each user’s screenshots are private personal data, so each user is a tenant:

• Tenant = one of your users.
• Dataset = a device or capture source within that user, e.g. iphone, mac, ipad — or a single screenshots dataset with device and captured_at fields if you prefer to filter rather than partition.
• Search = content_filter: "images" with semantic: true. A query like “coffee shop meeting with Sarah” is encoded into image-embedding space and matches screenshots of that scene, even when they contain no readable text.

When a user signs up you create their tenant; when they delete their account you tear it down.

​

Worked example: document search for companies

Say you offer document search to companies. Each company’s documents must be isolated, so each company is a tenant:

• Tenant = one of your customer companies.
• Dataset = a content type within that company, e.g. pages, comments, attachments.
• Search = full-text, SQL, or semantic: true over one or more of those datasets.

​

When to deviate

A single shared tenant with a dataset-per-customer is only acceptable when the data is low-sensitivity and you are willing to own isolation in your own application layer (and accept that a bug there can cross customers). For anything personal, confidential, or regulated — screenshots, documents, health or financial data — use tenant-per-customer. It is the design Splendor is built around.

​

Next