Choosing an AI provider and model for Iris

Iris’s features inside rConfig (config analysis, drift narration, compliance narration, the optimisation assistant) run on one AI provider that you choose. It’s a decision about where your configs go, how good the answers are, and who pays. This page compares the options; the provider setup page has the steps.

The three questions

Where may configs go? Even with secrets obfuscated, a config describes your network in detail. Some organisations can send it to a cloud API under a data agreement; some need it to stay in their cloud tenant; some need it to stay on site.
How long are your configs? Config analysis sends the whole config on the first chat message, up to about 100,000 characters. Long firewall and core router configs need a capable model with a large context window.
Who pays, and how? Cloud providers bill per token. A self-hosted model costs hardware and power, and nothing per question.

Options compared

Option	Providers	Configs go to	Answer quality on long configs	Cost model	Design
Public API	Anthropic, OpenAI, Google Gemini (from V8.4.2 also Mistral, xAI, DeepSeek, Groq, Cohere, OpenRouter)	The provider, under its data terms	Highest	Per token	Public API
Private cloud	Azure OpenAI (your deployment)	Your Azure tenant and region	High, depends on the model you deploy	Per token, on your Azure bill	Azure OpenAI
Self-hosted	Ollama, or vLLM, LM Studio and LiteLLM through OpenAI-compatible	Nowhere outside your network	Depends on model size and hardware	Hardware	Self-hosted
Split	A self-hosted model for Iris, cloud agents over MCP for engineers who need more	Iris: on site. Agents: the agent’s provider	Mixed	Both	Split

rConfig’s Privacy tab classifies whatever you pick. Ollama or OpenAI-compatible on localhost, a private address, or a .local or .internal host counts as On-Premises. Named cloud providers count as Cloud Provider. An OpenAI-compatible URL anywhere else counts as External API Endpoint. See where your data goes.

Choosing a model

For config analysis, use the strongest model you can. Reading a 3,000-line firewall config and spotting the one rule that shadows another is hard. Small models get the easy questions right and the hard ones confidently wrong.
Context window matters as much as model size. A first chat message can carry the config, rConfig’s background knowledge (up to 14,000 characters) and your question. If the model’s context window is too small, the request fails or is silently cut. For self-hosted models see Sizing a local model.
Answer length is capped twice. Each feature has its own ceiling (4,096 output tokens for config analysis and drift narration, 2,048 for compliance narration and rule generation). The Max Output Tokens setting can lower that, never raise it.
Lower temperature suits config work. Focused, repeatable answers are what you want. Leave temperature at the default or lower it.

Switching later

You can change the active provider at any time in AI Center → AI Settings. Many teams start on a public API to see what Iris can do, then move to Azure OpenAI or a self-hosted model once they know which features they use. The features and prompts stay the same; only answer quality and cost move.

Frequently asked questions

Which AI providers does rConfig support?

Anthropic, OpenAI, Google Gemini, Azure OpenAI, Ollama, and any server that speaks the OpenAI chat API (vLLM, LM Studio, LiteLLM). From V8.4.2 Mistral, xAI, DeepSeek, Groq, Cohere and OpenRouter are also offered as providers.

Can rConfig's AI features run without sending configs to the cloud?

Yes. With Ollama or a self-hosted OpenAI-compatible server on your network, rConfig's Privacy tab classifies the provider as on-premises and configs never leave your environment.

Does the provider I pick in rConfig affect Claude Code or Cursor?

No. The provider setting only drives the Iris features inside rConfig. Agents connected over MCP use their own model.