Real answers, or an honest "I don't know".
Ask a model something your documents do not cover, and it often guesses. The Gateway answers from the sources you approve and cites them. It also carries your own procedures, the steps your organization follows, into every request that needs them. When no approved source covers the question, it says so.
- request
- req_2c81d0e4
- question
- outside every approved source
- answer
- fixed reply: no approved source covers this
- model call
- none
- spend
- $0.00
Better answers for less
- Pay less per answer. Repeated prompt prefixes bill at your provider's cached rate, and history is trimmed only where the trim keeps that discount. Every token added or removed is priced on its own line.
- See the source. Each grounded reply names the passages it used from your approved sources.
- Your procedures, applied, not just read. A procedure your team wrote, like an escalation runbook or a support policy, is enacted whole in the request, not chopped into fragments. The model follows the steps you wrote, the same way every time.
- No guessing. A question outside every approved source gets a fixed reply. It makes no model call and costs nothing.
- Long chats stay sharp. Long threads are held to a token budget instead of growing without limit, and the exchanges that matter stay in.
Objections
- "A cheaper model will quietly get worse." Grounding and procedures do not depend on model size. A smaller model given your sources and your steps often answers as well as a larger one guessing without them.
- "Our data cannot leave our control." The sources and procedures the Gateway reads from are your own, scoped to the workspaces you grant. A request outside those grants cannot reach them.
- "We already have a RAG tool on top of our app." That tool grounds its own app's traffic. This grounds every client, every team, every model, behind the one endpoint you already changed to.
- "Retrieval tools make things up anyway." A question outside every approved source gets a fixed refusal, not a guess, and that refusal makes no model call and costs nothing. When your slot opens, test it yourself.
Try to make it guess
When your slot opens, ask questions your approved sources do not cover.
- Read the replies. Each one says it cannot answer. None guesses.
- Check the records. No model call was made, and the spend is zero.
Test it on your own documents.
We admit customers in cohorts, in waitlist order. Don't miss the next one. When your slot comes up, you start on one real workload, on your plan from day one. You keep a report on every grounded answer and every refusal. If we are not earning our fee, leaving is the same URL change, reversed.