AI jargon, decoded

An LLM running in a loop with tools, deciding its own steps to reach a goal.

Software around a model that provides the agent loop, tools, memory, permissions, scheduling and interfaces — e.g. Hermes Agent, Claude Code.

Artificial General Intelligence — a hypothetical AI matching humans across most tasks. Not needed to get business value today.

Application Programming Interface — a way for software to talk to other software, e.g. sending prompts to a model programmatically.

A secret token identifying your account to an API. Never paste keys into prompts or public files.

Splitting documents into smaller passages before embedding them for retrieval.

The maximum amount of text (in tokens) a model can consider at once — its working memory.

A measure of how similar two embedding vectors are; used to rank search results.

Machine learning with large, many-layered neural networks. Powers modern vision, speech and language AI.

A list of numbers representing the meaning of text; similar meanings produce similar vectors.

Test sets used to measure whether an AI system gives correct, safe outputs — re-run after every change.

Including a few input→output examples in a prompt to demonstrate the desired behavior.

Further training a model on your own examples to change its style or behavior. Usually not the first choice for adding knowledge.

When a model outputs a structured request to run a named tool with arguments; your code executes it. Also called tool use.

AI that creates new content: text, images, audio, video or code.

A file format for quantized models used by llama.cpp, Ollama and LM Studio.

Giving the model source material and instructing it to answer only from it, reducing hallucination.

Limits and checks around AI behavior: permissions, step limits, validation, human approval.

A confident but false or fabricated model output.

A design where a person reviews or approves AI outputs or actions before they take effect.

A property where running the same step twice has the same effect as once — prevents duplicates when triggers repeat.

Using a trained model to generate outputs. What happens every time you send a prompt.

A simple text format for structured data. The standard way AI output is passed into automations.

A description of the exact structure JSON must follow; used for tool parameters and structured outputs.

Large Language Model — a neural network trained on huge amounts of text to predict the next token.

A model running on your own hardware, so data never leaves your machine.

Systems that learn patterns from data instead of following hand-written rules.

Model Context Protocol — the open standard for connecting AI apps to tools and data through reusable servers.

A program that exposes tools, resources and prompts to MCP-compatible AI apps.

An Ollama recipe combining a base model, system prompt and parameters into a named custom model.

Several agents with different roles or tools collaborating on a task.

Models that handle more than text — images, audio or video.

A fair-code workflow automation platform with native AI agent nodes; free to self-host.

A free tool to download and run open-weight models locally with one command and a local API.

A self-hosted, multi-user ChatGPT-style interface for Ollama and other APIs.

A model whose weights are published so you can download and run it (Llama, Qwen, Gemma, Mistral, gpt-oss…).

An API that follows OpenAI's request format, so tools built for it work with other providers — including Ollama at /v1.

A pattern where a lead model splits a task and delegates sub-tasks to worker agents.

The learned numbers inside a model; model size is measured in them (e.g. 8B = 8 billion).

The instruction and input you give a model.

Splitting a task into sequential prompts where each output feeds the next.

Malicious instructions hidden in content an AI reads, trying to hijack its behavior or tools.

Compressing model weights (e.g. to 4-bit) so models need less memory, with small quality loss.

Retrieval-Augmented Generation — retrieving relevant documents and giving them to the model before it answers.

Using a second model to reorder retrieved passages by relevance before generating an answer.

A model that spends extra tokens thinking step by step before answering; better at complex problems.

Classifying an input and sending it to a specialized prompt, model or workflow.

A saved, reusable set of instructions an agent can load for a specific kind of task.

Model output constrained to a defined format, typically JSON matching a schema.

Persistent instructions that define a model's role, rules and behavior across a conversation.

Setting that controls randomness. Low for facts and extraction; higher for creative work.

A chunk of text (~¾ of an English word). Models read, write and bill in tokens.

A function an agent can call — search, read a file, send a message, query a database.

The date after which a model has no built-in knowledge.

The event that starts a workflow — a form submission, new email, schedule or webhook.

A database optimized to store embeddings and find similar ones quickly (pgvector, Chroma, Qdrant).

A URL that receives data from another system and starts a workflow.

A predefined sequence of steps triggered by an event; may include AI steps.