Acme RAG
Open the console
Live interactive demo

Enterprise RAG with access control built in

A multi-agent document assistant that enforces who can see what, cites every claim, and refuses rather than hallucinates. Try it as three different roles in under a minute.

Open the assistant

Try asking

Ask the first one as a Viewer and you are refused. Switch to Admin in the sidebar and ask again.

What makes it enterprise-grade

Access control at the data layer

Switch role and watch the same question get answered for an admin and refused for a viewer. The filter runs during retrieval, so a restricted passage never reaches the model.

Multi-agent, grounded answers

Each query runs visible agent steps - understand, retrieve, compose, answer - and every reply cites the passage it used. No citation, no claim.

Guardrails that refuse and redact

Injection attempts are blocked before the model runs, citations are validated against what was retrieved, and PII can be masked out of answers.

Operator controls, live

Swap the model, bring your own key, and flip guardrails without a redeploy. Re-tier a document and watch access cascade to every chunk of it.

Retrieval runs with no setup

Access control, retrieval and the guardrails work out of the box. To have a model write the answers, add your own API key: OpenAI, Anthropic, Google, Groq, OpenRouter and more. It is held in an httpOnly cookie in your browser and sent only to the provider you pick.

Add a key

Retrieval you can put in front of an auditor

A multi-agent assistant over a small company knowledge base. Unlike a general chatbot it enforces who may read what, cites the passage behind every claim, and refuses when the evidence is not there. Three things make that true.