Glossary

Admin panel The browser-based management surface at admin.nufi.me. Where admins configure NUFI, manage users, roles, groups, and capability grants.

Agent A saved AI assistant with its own name, instructions, tools, and knowledge files. NUFI's equivalent of a "custom GPT".

Backend The actual AI provider behind a model — OpenAI, Anthropic, Google, a self-hosted vLLM server, etc. The gateway routes to it.

backend_type A required label on every registered model — gpu, npu, or cloud. Used by reporting and the canary slider.

Bootstrap The setup script that fills in .env secrets, pulls images, and starts the stack on first install. Idempotent — re-running is safe.

Budget Spend cap, in USD, on a user or API key. Refreshes on the configured interval (default 30 days).

Capability The unit of authorisation inside the admin panel. Constants like MANAGE_USERS, MANAGE_CONFIG, IMPERSONATE. Roles bundle capabilities; users inherit capabilities via roles.

Chat The conversational surface end users interact with — at chat.nufi.me in production. The default NUFI experience.

Console The self-service developer UI at console.nufi.me. Where users generate API keys, watch their usage, and tune their per-key limits.

Conversation A single chat thread, with one or more messages. Stored in NUFI's database under the user's account.

Endpoint A named integration with one or more models. The default NUFI deployment exposes its models through one endpoint named "Nufi".

Gateway The single chokepoint every AI request goes through. Handles routing, authentication, budgets, rate limits, and trace emission. Reachable at api.nufi.me for API users.

hardware_id A free-form label on every registered model identifying which piece of hardware serves it — e.g. dgx-a100-01, together-cloud, mac-local. Reports aggregate by this.

Override A scoped config value that beats the base config for a specific role, group, or user. Resolved by priority — see Scoped overrides.

Preset A saved bundle of {model, parameters, system prompt} with a name. Users start new conversations from a preset.

Profile (in the config editor) A scoped override on a configuration field — has a principalType (role / group / user) and a priority. The admin panel's "Field Profile Popover" lists every profile on a field, sorted by priority.

Resumable streams The feature where assistant replies automatically reconnect and resume if the user's connection drops. Backed by Redis. Required for multi-instance chat deployments.

Safety filters The pre-call hooks the gateway runs on every prompt — personal data detection and prompt-injection scanning. Block or mask the prompt before it reaches the AI provider.

SKILL.md A reusable instruction bundle that can be attached to an agent or called manually via /skill-name. A lighter-weight unit than an agent.

Trace A top-level event in the trace viewer (usually one chat completion). Inside a trace, observations record nested calls (one per AI call, tool call, or guardrail decision).

Trace viewer The observability surface at langfuse.nufi.me. Records every AI request with the prompt, reply, cost, and labels.

Virtual key (API key) A NUFI API key issued to a specific user (or service account) with its own budget, rate limits, and model allow-list. Users self-issue them from the console.