How to configure AI capabilities and context in rConfig V8

AI Configuration is where you decide which AI features your team gets. Each of the four capabilities has its own switch, so you can roll out config explanations to everyone while holding back, say, compliance narration until your auditors have seen it. This page also explains how rConfig feeds each capability the right background knowledge so answers read like they came from a network engineer, not a general chatbot.

Prerequisites

• The aicenter.update permission. The AI Configuration tab is shown only to users whose role has it (Admins do). From V8.4.2 this follows the permission, not the role name.
• An active provider (see Set up an AI provider). You can toggle capabilities without one, but they won’t answer until a provider is active.

The four capabilities

Capability	Where users find it	What it’s good for	Sends device config?
Config Analysis & Explain	Ask AI in the config viewer, and the lightbulb menu in the editor	“What does this do?”, spotting misconfigurations, security reviews	Yes
Drift Detection & Narration	Analyze with AI on the Config Changes panel and in Change Pulse	Explaining a change during change review or an incident	Diff only
Policy & Compliance Narration	Narrate with AI on compliance results, AI Generate Rules in the policy editor	Explaining failures for a report, drafting new policy rules	No (optional paste for rule drafting)
Optimization Assistant	AI Assistant on Settings → System Optimization	Tuning PHP, the database, and Horizon queues	No

A sensible rollout order

Start with Config Analysis & Explain and Optimization Assistant: they’re the easiest to judge. Add Drift Detection & Narration once your team trusts the answers, and Policy & Compliance Narration once compliance owners have reviewed sample output.

Turn a capability on or off

1. Open AI Configuration

1. Click AI Center in the left sidebar.
2. Select the AI Configuration tab.

2. Flip the toggle

Each capability card has a toggle. The change saves and applies immediately, with no confirmation and no save button. Turning a capability off removes its button from every screen for every user; turning it on brings the buttons back on their next page load.

The card also shows Current model (always “Provider default” in this release), Usage this month (a dash, reserved for a future release), and a disabled Override model button. Per-capability models aren’t available yet; every capability uses the active provider’s model.

How rConfig grounds AI responses

A general-purpose model doesn’t know how rConfig works, your naming conventions, or what a good config looks like in your network. rConfig adds that context to each request before it reaches the model, in one of two ways.

The context harness (config-work capabilities)

Config Analysis, Drift Detection, and Compliance Narration get a block of networking domain knowledge added to every request: how vendors structure configs, what common security findings look like, and how to read a diff. From V8.4.2, rConfig loads this built-in knowledge automatically when you install or upgrade. On top of that sits a customer layer: your own conventions and standards, if you’ve added any (see below).

Each capability has its own context budget, so requests stay fast and work on smaller local models:

Capability	Context budget
Config Analysis & Explain	14,000 characters
Drift Detection & Narration	10,000 characters
Policy & Compliance Narration	10,000 characters
Optimization Assistant	8,000 characters

This is a cap on background knowledge per request, not a spend limit.

The knowledge layer (Optimization Assistant)

The Optimization Assistant doesn’t need networking knowledge; it needs to know how rConfig itself is tuned. For each question it searches a local index of rConfig documentation (PHP settings, database tuning, Horizon, scheduled tasks) and adds only the most relevant passages. The index lives in rConfig’s database, so answering a question doesn’t need internet access. See Optimization Assistant for how to refresh it.

Teach the AI your conventions

Customer knowledge entries let you tell the context harness things only your team knows. Each entry has a title, up to 10,000 characters of content, and one of four categories:

Category	Example
Convention	“Loopback0 is always the router ID and the BGP update source.”
Golden Config	Your standard AAA, NTP, and logging block for access switches.
Naming Standard	“Hostnames follow <site>-<role>-<nn>, for example dub-core-01. Roles: core, dist, acc, fw, wan.”
Design Principle	“No device in the DMZ may have SNMP v1 or v2c enabled.”

Active entries are added to every Config Analysis, Drift, and Compliance request, after rConfig’s own knowledge. From V8.4.2, at least 40% of each capability’s budget is kept for your entries (5,600 characters for Config Analysis, 4,000 for Drift and Compliance), and they can use more when rConfig’s own knowledge needs less. Entries are added in category order, then oldest first. An entry that doesn’t fit in the space left is left out whole rather than cut off mid-sentence, so keep each entry short and specific.

No form in this release

There’s no form for customer knowledge in the AI Center yet. The AI Center endpoints behind it (/api/ai/knowledge/customer) accept a signed-in rConfig browser session with the aicenter.update permission, not REST API or MCP tokens. Until a form ships, add an entry from the rConfig server:

php artisan tinker --execute '\App\Models\CustomerKnowledgeEntry::create(["title" => "Hostname convention", "category" => "naming_standard", "content" => "Hostnames follow <site>-<role>-<nn>, for example dub-core-01."]);'

category is one of convention, golden_config, naming_standard or design_principle. Set is_active to false on an entry to stop using it without deleting it. rConfig’s built-in knowledge can be read but not changed.

Common gotchas

Common configuration mistakes

• Turning a capability off doesn’t stop provider billing: it only hides the feature. Control spend in your provider’s console.
• All capabilities share one provider: switching the active provider on AI Settings changes it for every capability.
• “Enabled” doesn’t mean “working”: the toggle controls whether buttons appear, not whether a provider is ready. If an enabled capability answers with No AI provider is activated, activate a provider.
• Users don’t see a newly enabled capability: they need to reload the page, and they need to have accepted the AI Center terms.

How the harness works

The context harness explains what Iris adds to each request, how the budget is shared, and how to write knowledge entries that work.

What’s next

• Use Config Analysis & Explain in the config viewer
• Review what data each capability sends to your provider
• Fix a capability that isn’t working