---
title: "AssistantGeneral Whitepaper"
description: "A technical reference for security review and procurement: deployment model, inference layer, agent loop, retrieval, governance and compliance posture."
url: "https://royalsoftworks.com/products/assistant-general/whitepaper/"
source: "https://royalsoftworks.com"
format: "markdown"
note: "Markdown rendering of the HTML page at `url`. Same content, same canonical URL."
---

Whitepaper

# Architecture & Governance: AssistantGeneral

A technical reference for security review, procurement and IT architecture teams evaluating AssistantGeneral for deployment.

Private beta. Builds go to testers directly; ask to join. [Join the beta](https://royalsoftworks.com/contact/?product=assistant-general&tier=beta)

Contents

1. 01Executive summary
2. 02Deployment model
3. 03Inference layer
4. 04The agent execution loop
5. 05Profession workspaces and practice packs
6. 06Retrieval architecture
7. 07Data connectors and external action
8. 08Document generation
9. 09Enterprise governance
10. 10Data protection
11. 11Compliance posture

01

## Executive summary

AssistantGeneral is a local-first agentic AI platform that runs on standard desktop hardware (Windows today; Linux to follow) and, for organisations, ships with a complete governance control plane. The platform is designed around one constraint: capability should never require sending data off the customer's infrastructure.

This document describes the inference architecture, the agent execution loop, the retrieval system, the profession workspaces, the data connector model, and the enterprise governance layer — in enough detail to support a security review and a deployment plan. The platform is in a private beta; the changelog lists each build and what it does not yet do.

02

## Deployment model

AssistantGeneral installs as a desktop application with an embedded local server process. There is no mandatory external service: inference, retrieval, and tool execution all happen on the host machine. Cloud model providers are integrated as opt-in routes selected per request, never as a default dependency.

- Single-user installs run standalone once the engine and model have downloaded on first run — no network calls required after that
- Enterprise and Government editions add a self-hosted admin control plane on infrastructure you operate
- No Royal SoftWorks-operated service sits in the runtime data path

03

## Inference layer

The default inference engine is llama.cpp, running any GGUF-format model fully offline. This covers Llama, Mistral, Qwen, Phi, Gemma, DeepSeek and most open-weight model families. For tasks that benefit from a larger frontier model, the platform supports one-click routing to nine cloud providers — Anthropic, OpenAI, Google Gemini, xAI, Groq, Mistral, DeepSeek, Together AI and OpenRouter — or any OpenAI-compatible endpoint.

Multiple local machines or GPUs can be pooled into a single routed cluster, distributing inference load across available hardware without any single point depending on external compute.

04

## The agent execution loop

Rather than a single forward pass, every request that needs work runs through a structured pipeline: plan, retrieve, act, verify, finish. A first, lightweight pass answers simple conversational messages directly; when the user's documents hold the answer, a second fast path answers from the retrieved passages in a single call, with citations, and escalates if they do not suffice. Everything else goes to the pipeline. The planning stage decomposes the request into steps before any tool runs. Retrieval pulls grounding context from the knowledge base. Acting invokes tools — file operations, document generation, connector queries, workflow steps — under policy control. A groundedness gate checks the draft's specific claims against what was retrieved before the response is finalised: several unsupported claims trigger a targeted re-retrieval and one revision, a stray one is flagged for the answer to hedge. Generated files are re-opened and audited for errors, placeholders and broken references before they reach the user; a self-review pass over written answers is built but switched off in the beta.

This loop is the basis for the platform's reliability claims: outputs are gated on verification rather than produced in a single uncontrolled generation step.

Two stages are conditional. When a request has independent parts, the model may delegate each to a helper: a separate run of the same loop with an empty transcript, its own context window, a role-specific tool scope and its own groundedness gate, returning one bounded report. Helpers cannot delegate further, and concurrency is measured from the inference engine's actual slot capacity with one slot always reserved for the parent conversation. For legal work, an optional citation check sits after the groundedness gate: it extracts the draft's citations and hands each to a per-jurisdiction verifier. Only an affirmative negative from a source that answered sends the draft back; a timeout or a gap in coverage never condemns a citation, and remote text reaches the model only as sanitised, fenced, passive data.

05

## Profession workspaces and practice packs

A chat can be bound to a profession — law office, judiciary, hospital, clinic, HR or recruitment — and then opens in a workspace composed for that job over the same backend: an object the work is organised around, a scope limiting which document collections the chat may retrieve from (enforced in retrieval, not in the interface), a set of everyday actions, an evidence rail, and the document the job produces. Several workspaces can be open at once as tabs. A workspace stamps defaults onto a chat when it is created — persona, scope, citation check — and never steers a running request; after creation they are the chat's own settings.

A practice pack adds the jurisdiction: the authorities normally cited and their citation format, the usual procedures as staged timelines, the documents normally produced with their required sections, and professional checklists. Twenty-four packs ship, covering the six professions in an international baseline, Serbia, Germany and the United States (federal law only). Packs drive six views — comparison, extraction, cited answers, procedure timeline, drafting and checklist — which run through the ordinary chat path and therefore inherit its scope, persona and audit. The procedure view's deadline calculator is local date arithmetic and never consults a model.

Judiciary, HR and recruitment are high-risk uses under Annex III of the EU AI Act. Their workspaces carry a non-dismissible notice that a person decides, and administrators control which workspaces and packs each group receives through the same policy engine as every other capability.

06

## Retrieval architecture

Documents are indexed locally into a LanceDB vector store. Retrieval is hybrid: dense vector similarity and sparse keyword search are computed independently and fused using Reciprocal Rank Fusion, then the merged candidate set is reranked with a cross-encoder and diversified with Maximal Marginal Relevance to reduce redundant chunks. Recency weighting biases toward newer documents when relevant.

Ingestion supports PDF (scanned pages are read by the loaded vision model rather than a separate OCR engine), Word, Excel, PowerPoint, CSV, plain text, Markdown, HTML, saved email files, audio recordings (transcribed locally), and YouTube transcripts. Beyond flat retrieval, a knowledge graph links documents that share topics, tags, categories and keywords, so retrieval can follow those links to related material instead of stopping at a single passage. Organisations can maintain multiple separate knowledge bases with independent access policies.

07

## Data connectors and external action

Twenty built-in connectors synchronise from common enterprise sources: AWS S3, Azure Blob Storage, Google Cloud Storage, Azure DevOps, SharePoint, OneNote, OneDrive, Outlook, Google Drive, Gmail, Google Calendar, Google Contacts, IMAP email, Dropbox, PostgreSQL, MySQL, BigQuery, Snowflake, a configurable web crawler and a custom connector. Database connectors enforce read-only query validation before execution.

Beyond ingestion, Model Context Protocol (MCP) support allows the agent to both read from and act within nine connected accounts — Slack, Microsoft Teams, Discord, Notion, Jira and Confluence, GitHub, GitLab, Linear and HubSpot — under the same policy engine that governs every other tool, which can limit them per connector and per tool. The platform can also run as an MCP server, exposing its own capabilities to other MCP-compatible clients, and as an MCP client connecting to custom internal servers.

08

## Document generation

Office document generation runs inside a sandboxed Python (Pyodide/WASM) execution environment with no filesystem access beyond its working sandbox; any network request it makes goes through the same egress gate as the rest of the application, so Offline mode or an administrator's egress policy closes it. Generated .xlsx, .docx and .pptx files are produced with real, computed values — not descriptions of a document. After generation, the platform re-opens the file, audits it for formula errors, placeholder text and broken references, and applies automatic repair before returning it to the user.

09

## Enterprise governance

Enterprise and Government editions add a control plane: an administrator console, identity integration, and centrally enforced policy. A granular role-based access policy is pushed to every managed client and enforced locally on that client — controlling which models, tools, workflows, connectors, data sources, workspaces, practice packs, helpers and UI surfaces each group may use. Policy normalisation is fail-closed, and policy records carry a schema version so that a capability added in a later release is not silently denied to a policy written before it existed. Each desk stores the policy it was given and applies it at every start, so it stays in force when the console is unreachable; a stored policy that has been altered or deleted locks the desk to deny-all. Action limits apply per connector, per MCP tool and per website. Anything an administrator does on a desk — running a tool, moving a file, changing settings — is recorded there and shown to its user, and a read-only auditor role can review the audit without changing anything.

Identity integration uses OIDC single sign-on with SCIM for automated user and group provisioning. Both are implemented on the admin server and, during the beta, are enabled per organisation as part of a pilot while the self-service configuration screen is completed. SAML 2.0 is not offered until XML signature validation is wired in; the endpoint fails closed rather than accept an unverified assertion.

Every governed action is recorded to a fleet-wide, append-only audit log in which each entry is chained to its predecessor, aggregated at the admin console, giving compliance and security teams a single point of visibility across every deployed instance without any of the underlying data leaving customer infrastructure.

10

## Data protection

Encryption at rest is on by default. Configuration, chat sessions and task notes are encrypted with AES-256-GCM under a key wrapped by the operating system's credential store (DPAPI on Windows). The knowledge base — document databases, the vector and full-text index, the embedding cache — is placed under Windows EFS, because the index engine has no encryption of its own; where EFS is unavailable the application reports whether BitLocker covers the volume, or that the knowledge base is unencrypted. Traffic between managed desktops and the admin server is TLS-only: the server presents a self-signed certificate that each desktop pins on first contact and verifies on every connection. An optional PIN lock protects the application on shared or unattended machines and is enforced in the main process, which withholds configuration and data until it is satisfied. Scheduled backups are encrypted with AES-256-GCM under a key derived from an operator-held passphrase rather than the machine's keystore, so an archive remains recoverable after the originating machine is lost. Remote web pages are rendered in an isolated, sandboxed window with its own session and no access to the application. Tool execution passes through a three-tier permission model and a sandboxed working directory, so file and shell operations are scoped and confirmable rather than unrestricted.

There is no telemetry and no application log: no background process reports usage data externally. The application's own outbound requests are an update check after launch, licence revalidation for paid licences, and a lookup of available engine builds; Offline mode or an administrator's egress policy stops all of them. The only diagnostic artefact is a local crash file containing a timestamp, the version and the error with its stack trace. It is not filtered, so an error that quoted part of a request would appear in it, but it never leaves the machine unless the user sends it.

11

## Compliance posture

Because data processed locally never transits Royal SoftWorks infrastructure, AssistantGeneral does not establish a data-processor relationship under GDPR for that processing, and no Data Processing Agreement with us is required for locally-run deployments; the customer remains the controller of what it processes, and HR and recruitment uses still need the customer's own impact assessment. The architecture suits network-restricted environments: once the engine and model are in place after first-run setup and a paid licence has been activated — which needs a connection once, as there is no offline activation file yet — every core capability (inference, retrieval, tool execution, workflows and voice) operates without an outbound network connection, and Offline mode closes every outbound path the application has. The Government licence pairs this posture with the admin control plane, run inside the customer's own network.

Companion papers

## Where this goes deeper.

This document is the platform reference. Four shorter papers take a single dimension of it further — one architectural, three by industry.

[### The Private AI Stack How reasoning, retrieval, embeddings, speech and document generation all run on the user's own device by default — and why that is a structural property rather than a setting.](https://royalsoftworks.com/resources/private-first-architecture/)

[### Governing AI Without a Vendor Cloud The identity, policy and audit plane described above, written for enterprise and government IT: RBAC enforced on every client, company sign-in in pilots, and running with no internet once set up.](https://royalsoftworks.com/resources/enterprise-ai-governance/)

[### Private Legal AI Privilege and confidentiality, contract review, case-law research, deadlines and drafting in the law office and judiciary workspaces — without client data leaving the firm's machines.](https://royalsoftworks.com/resources/legal-ai-on-device/)

[### Private Clinical AI Clinical documentation and literature research tied to primary sources, with a groundedness gate on claims — not for diagnosis, triage or treatment decisions.](https://royalsoftworks.com/resources/healthcare-ai-on-device/)

[### Private Finance AI Validated spreadsheets whose figures tie out, and read-only analysis of data you would never paste into a cloud tool.](https://royalsoftworks.com/resources/finance-ai-on-device/)

## Request the full security questionnaire response.

We respond to vendor security reviews, SOC-style questionnaires and architecture diagrams on request — handled directly by the engineering team.

[Request the questionnaire](https://royalsoftworks.com/contact/?product=assistant-general&intent=questionnaire)

Private beta. Builds go to testers directly; ask to join. · [Join the beta →](https://royalsoftworks.com/contact/?product=assistant-general&tier=beta)

<!-- verse:royalsoftworks
     Luki 11:9, 10
     I ja vama kažem:ištite i daće vam se:tražite i naći ćete; kucajte i otvoriće vam se. Jer svaki koji ište, prima; i koji traži, nalazi; i koji kuca, otvara mu se.
     https://www.wordproject.org/bibles/verses/serbian/02_prayer.htm
-->
