Add AI Era Personalisation with a White-Label RAG API

An API that lets your product know each of its customers and personalise to them, using retrieval-augmented generation over that customer's own content, under your brand, without you building the retrieval. Context Link's RAG White-Label plan holds one private vector index per end-customer inside your account, built from their website and/or the data you feed it. Your backend provisions an index at signup, pushes or crawls that customer's content, and calls two endpoints with that customer's token: GET /api/v1/context for the matching passages as markdown, and GET /api/v1/question for a concise grounded answer with citations. Everything is server-to-server. Your customers never see Context Link.

Three things, in practice. AI responses: anything your product generates for a customer can be grounded in their own material first, so an outline, a draft, a recommendation or an answer reflects what they actually sell and say. Onboarding: read the account's top_level_context paragraph and the first screens a new customer sees can already speak to their business rather than a placeholder. UI: defaults, examples, suggested prompts and empty states can be filled from the customer's own content instead of generic copy. All of it comes from one index per customer that you never have to host or tune.

Every end-customer is an API account with its own token, and every passage in the index carries that account's id. Retrieval with a customer token filters to that account plus anything you deliberately publish at the organisation level, such as your own help docs. Content pushed with a customer's token is only ever visible to that customer, and there is no scope parameter to get wrong. Your provisioning token is the one credential that can search across every account, which is why it belongs in your core backend only.

Two things, in any combination. Their public website, if you give a site URL at provisioning: Context Link crawls it (capped at 500 pages), re-syncs it on a schedule, and fires a signed webhook when the first crawl is done. And anything your backend can send as markdown through the Custom Connections API, by a stable id: dashboards, reports, help articles, notes. You never store or index it yourself; that is half the point. Accounts with no website are fine; they are queryable as soon as content is pushed.

The parts of RAG that take the time. A crawler that gets past bot defences and keeps the index current. Chunking at 800 to 3,200 characters with boundaries that keep meaning intact. Standard OpenAI embeddings in a hosted vector database, re-run only when content actually changes, and not a proprietary encoding you would be locked into. A retrieval pipeline that goes beyond a raw vector lookup: a blended rerank weighting similarity by passage length and source weight, fuzzy boosting on titles and tags, a filter that drops boilerplate, and expansion to the full document when a query lands squarely on one page. Named Modes let you weight sources differently per use case. And the hosting, backups, and per-tenant isolation of all of it.

Call GET /api/v1/context when your own model will do the reasoning: it returns the matching passages as markdown, which is rich input for a generation step or an agent. Call GET /api/v1/question when you want to show an answer directly: one short paragraph plus numbered citations, which also protects an agent's context window instead of dumping the customer's whole corpus into it. There is a third, cheaper signpost that needs no query at all: every account carries a top_level_context paragraph describing who that customer is, written automatically once their index has content.

$50 a month per bundle of 10 end-customer accounts, with a seven-day free trial. When you provision an eleventh customer, a bundle is added automatically. There are no token or usage fees on top: /context is retrieval only, and each bundle carries 5,000 concise-answer requests a month. Bundles never shrink on their own; you downgrade explicitly if your fleet does.

No. Context Link does not ship an embeddable chat surface, and it does not expose MCP to your end customers. It is the retrieval layer behind whatever you build: a help panel, an outlining feature, a search bar, an agent. You own the UI, the prompt, and the model. If a ready-made docs chatbot is the whole requirement, a tool like Kapa.ai is purpose-built for that.

OpenAI's file search and a vector database both give you the storage-and-similarity part. You still build ingestion, crawling, chunking, re-sync, reranking, per-tenant isolation, and the operational side of running it for many customers. The RAG White-Label API packages that whole layer with a fixed price per customer account, and returns markdown or a cited answer that works with any model you already use.

500 crawled pages per customer website, and crawling is deliberately slow to avoid being blocked, so wire up the index-ready webhook rather than blocking a request on it. Query rate limits are per customer token (5 per 10 seconds on /context, 2 per 10 seconds on /question), so one busy customer never starves another. Write rate limits are shared across your organisation at 5 per 10 seconds, so bulk pushes serialise. Custom connections hold up to 1,000 posts and 100 MB each. The full tables are in the errors and limits reference.