---
title: "How AssistantGeneral Works"
description: "AssistantGeneral runs a real loop — plan, find, do, check, review, finish — grounded in your own documents and verified before it answers. All locally."
url: "https://royalsoftworks.com/products/assistant-general/how-it-works/"
source: "https://royalsoftworks.com"
format: "markdown"
note: "Markdown rendering of the HTML page at `url`. Same content, same canonical URL."
---

How it works

# It works the way a careful person would.

A chatbot writes the first thing that sounds right. AssistantGeneral plans the job, digs up the facts, does the work, then checks itself before it says a word.

02 The engine

## It doesn't guess. It works.

A chatbot writes the first thing that sounds right. AssistantGeneral runs a real loop: it plans, gathers, acts, and checks itself, the same way a careful person would.

1 ### Plan It works out the steps before it does anything, like a person sketching an approach.

2 ### Find It pulls the facts it needs from your own documents, right there on your machine.

3 ### Do It uses tools: builds the file, runs the numbers, queries a database, or calls an app.

4 ### Check It makes sure every claim is actually backed by a real source before it trusts it.

5 ### Review It reads its own work, catches mistakes, and fixes them.

6 ### Finish It hands you a clear answer or a finished file.

For your IT team

- **Loop** — Route → plan → retrieve → tool-call → groundedness gate (closed re-RAG) → review → finish
- **Safety** — Loop / runaway protection; durable cross-session task memory; risky tools always confirm
- **Personas** — Built-in experts (Lawyer, Doctor, Cook, Mechanic, Programmer) plus custom agents and skill packs
- **Transparency** — Live plan tracker, activity log, and an inference trace lens for every step and tool call

Why it's trustworthy

## Four habits that keep it honest.

Each step exists to stop one of the things that makes people distrust AI in the first place.

### Answers from your files

It retrieves the facts it needs from your own documents and points to the exact source it used — no made-up references.

### Does the work with real tools

It builds files, runs numbers, queries databases and calls apps, rather than describing what you could do.

### Checks before it trusts

A groundedness gate confirms every claim is backed by a real source; a review pass catches and fixes its own mistakes.

### Shows its working

A live plan tracker, an activity log and an inference trace let you see every step and tool call it made.

Under the hood

## From your question to a finished answer.

A technical walk-through of what actually happens between you asking and the agent replying.

1. 01

   ### Route

   The request is classified and routed to the right model — local by default, or a cloud provider if that task is configured to use one.
2. 02

   ### Plan

   A planning pass decomposes the request into concrete steps before any tool runs, the same way a careful person sketches an approach first.
3. 03

   ### Retrieve

   Hybrid retrieval pulls grounding context from your knowledge base: vector similarity and keyword search are fused with Reciprocal Rank Fusion, then reranked and deduplicated.
4. 04

   ### Act

   Tools execute under policy: file operations, document generation, connector queries, workflow steps — each gated by the three-tier permission model.
5. 05

   ### Verify

   A groundedness gate checks every substantive claim against a retrieved source. Anything unsupported triggers a closed re-retrieval pass before it can reach the answer.
6. 06

   ### Review

   A review pass re-reads the full output, looking specifically for errors, placeholders and inconsistencies, and repairs them before finishing.
7. 07

   ### Finish

   You get the answer or the finished file — plus a full trace of every step, tool call and source it used to get there.

Where your data goes

## Everything happens inside the box.

Your files, the index built from them, the model that reads them and the answer it writes — all four live on the same machine. The arrow out is drawn so you can see it is not there.

Your documents and a local index feed the model, which writes your answer. All four sit inside your machine. A connection to the internet is drawn and marked as never used.

You can point it at a cloud model on purpose, per task, if you want one. It is a setting you turn on, not a default you have to find and turn off.

## And none of it phones home.

The whole loop runs on your machine by default. See exactly how your data is protected.

[Security & privacy](https://royalsoftworks.com/products/assistant-general/security/) [Read the technical whitepaper](https://royalsoftworks.com/products/assistant-general/whitepaper/)