About

A gate for the one question retrieval evals don't ask.

Relevance benchmarks tell you whether your retrieval is good. They say nothing about whether it is allowed. raggate exists to make that second question testable, repeatable, and documentable — with the same discipline a test suite gets.

Who it's for

Three people, one problem

The engineer shipping RAG

You are about to roll out an index change, an embedding upgrade or a new connector, and you would like to know what it exposes before your users find out. You want a command with an exit code, not a dashboard.

The security reviewer

You have been asked to sign off on an AI feature that reads from a corpus you do not control. You want evidence you can read months later, not a screenshot of a green tick.

The platform team with tenants

You run one index across many customers or business units, and "similar documents" is exactly the boundary you cannot afford to blur.

How it is built

Opinions the code actually holds

Offline and deterministic by default

The default embedder is seed-stable and runs locally, so the same inputs produce the same verdicts on any machine. Nondeterministic gate results are treated as a defect, not as variance to be tolerated.

Fail loudly, never silently clean

An unreachable backend, a malformed response or an unresolvable document produces an error, not an empty result set. A tool whose job is finding leaks must never report a clean run because it quietly retrieved nothing.

Evidence over assertion

Every finding is replayable from its record, and the source is Apache-2.0 so the logic behind a verdict can be read rather than trusted. Where the tool's coverage is narrower than a claim would suggest, it says so in its own output.

Reversible before destructive

Remedies are ordered safest-first: narrow retrieval before you move data, and move data before you delete any. Rollback artifacts are written before mutations, not after.

Where things stand

Early, and saying so

raggate is a young project. The detection path is well covered by tests and exercised on every change; the commercial path has had exactly one real purchase, which was our own end-to-end test of checkout.

There are no customer stories on this site because there are no customers to tell them yet. Nothing here implies adoption we do not have. When that changes it will be because someone agreed to be named, not because a marketing page needed filling.

What exists instead is the thing itself: a repository you can clone, a sample corpus with planted leaks, and output you can inspect line by line. That seemed like a better opening argument than adjectives.

Contact

Get in touch

Questions, a pilot conversation, a bug, or a disagreement about how the oracle should treat some edge case — all welcome at the same address.

WhatWhere
Anything at allhitliqiwei@gmail.com
Security disclosureSame address — see Security for what to include
Terms and data handlingTerms · Privacy