FAQ

Does NUFI host AI models?

No. NUFI routes through its gateway to whatever AI provider you point at — your own model server, OpenAI, Anthropic, Together, Groq, OpenRouter, anything that speaks the OpenAI Chat Completions API. Bring your own model.

Can I run only the chat?

Yes. Use the nufi-chat standalone deployment — chat application

• database, no gateway. Point it at any OpenAI-compatible endpoint.

Can I run multiple chat instances behind a load balancer?

Yes. NUFI supports horizontal scaling with Redis-backed resumable streams. Set the relevant Redis env vars. You also need a load balancer in front.

How do users sign up?

By default anyone with the chat URL can register. The first registered user becomes admin; the rest are regular users. In production, disable self-sign-up and create users from the admin panel (and/or wire in your identity provider).

What models do you recommend?

Depends on what you have. Common combinations:

• Local laptops — Ollama with qwen2.5:3b or llama3.2:3b.
• GPU pilot — a model server hosting qwen2.5-7b-instruct or llama-3.1-8b-instruct.
• Cloud — gpt-4o, claude-3.5-sonnet, gemini-1.5-pro for quality; gpt-4o-mini, claude-3.5-haiku for cost.

Mix and match — NUFI lets you expose all of them and per-user allow-list which ones each user can call.

How do I limit a user's spend?

Two layers:

1. Account budget — set on the user. Applies to everything the user does across all keys.
2. Per-key budget — set when the user issues a key. Tighter than the account budget.

Both are visible to the user in the console.

Can I expose only specific models to specific users?

Yes — three ways:

• Teams — bucket users by plan, attach an allowed_models list to each team.
• Per-user models — virtual keys can have an allowed_models set; you can edit per user from the gateway admin UI.
• Endpoint scoping — per-role / per-group config overrides on which endpoints and models are visible.

Pick the layer based on whether the gate is policy (gateway) or presentation (chat UI). Usually both — present only the models you have policy-allowed.

How do I add a new AI provider?

Use the add-model helper. Any OpenAI-compatible endpoint works — pass the base URL, the API key, and the model identifier. See Adding a model.

Where do my conversations live?

In NUFI's own database. They are not sent to AI providers beyond the prompt itself. If you want to delete a conversation, the chat UI does it. If you want to delete a user's entire history, the admin panel's "Delete user" cascade-deletes their conversations and files.

Is my prompt logged?

Yes — every chat completion is recorded with the full prompt and reply. This is the trade-off for being able to debug bad replies and answer "what did user X see?". If you have prompts you do not want logged, use temporary chat in the UI; temporary conversations are not persisted.

Can I disable conversation logging?

Yes. Ask your operator to remove the trace store credentials from the gateway and the recording stops. You lose observability and the console's usage analytics in exchange.

Why is the trace store growing so fast?

It stores every conversation in detail. At moderate usage, expect ~50 MB per 1000 conversations. Apply a retention policy if needed.

How do I sign people in via my identity provider?

NUFI supports OpenID Connect out of the box. Add the OIDC settings to your deployment, restart, and the sign-in screen gets a "Sign in with <provider>" button. The admin panel reads the same auth.

How do I rotate the gateway master key?

1. Generate a new key.
2. Set it in the deployment env, restart the gateway.
3. Update every consumer that uses it.
4. User-issued API keys keep working — only the master key changed.

Does NUFI work behind Cloudflare Access / Zscaler / corporate proxy?

Yes. NUFI is just HTTPS. Anything that fronts HTTPS works. If you put NUFI behind Cloudflare Access, signing in to Access lets you through to the chat sign-in.

Where do I report a bug?

The component that owns the surface:

• Chat behaviour or UI → the NUFI chat repo.
• Admin panel → the admin panel repo.
• Developer console → the console repo.
• Compose / env / install → the platform repo.

If you are not sure, file on the platform repo — the team will route it.