A RAG AI agent that admits what it does not know

This is an internal demo built by Neterius. It is not a client project and contains no third-party data: the corpus is our own public documentation.

What it does

• Answers questions about the loaded knowledge and cites the document each claim comes from.
• When retrieval returns insufficient context, it says so instead of improvising.
• Hands off to a person when the question exceeds what it can resolve.

How it is built

The knowledge is chunked and indexed with embeddings in a vector store. Each query retrieves the nearest fragments and the model answers only over that context. The handoff is a bounded agent action, not an on-the-fly model decision.

What the demo shows

• A valid query, with sources cited.
• A query with insufficient information, where the agent admits it.
• A handoff to a person, with a summary of the conversation.

Limits

An agent like this does not guarantee infallible answers. Quality depends on the knowledge loaded and on the evaluation done before it faces a customer. We publish no accuracy percentages or savings figures: without a comparable baseline they would mean nothing.