Modified

Plain language in a world of jargon

The AI world has its own language — and it shifts every quarter. This wiki explains what the terms actually mean, which programming language you should learn, and what every acronym really refers to.

Feel free to bookmark — handy when you come back to look something up.

Which track are you on?

Use agents
No code at all
n8n or Flowise — visual tools without a single line of code.
Build AI apps
Python
The most common language for AI, data, and automation. The most guides, the largest ecosystem.
Web & Frontend
TypeScript
AI frontends are almost always built in TS via Next.js or SvelteKit.

Most AI agent projects are built in Python or TypeScript. You do not need to learn both — pick one, AI fills in the rest.

Languages

Languages & Frameworks

Python
AI · Data · Automation

The standard language for AI and machine learning. Simple syntax, massive ecosystem. Perfect for agents, data analysis, and backend APIs.

Popular frameworks
FastAPILangChainLlamaIndexDjangoFlask
TypeScript
Web · API · Frontend

JavaScript with types — the same language, but with a layer that catches errors before the code runs. Dominant for modern web apps. Runs in the browser and on the server via Node.js.

Popular frameworks
Next.jsSvelteKitHonoExpressReact
Go
Backend · Microservices

Fast, simple, built for servers. Excellent for APIs and background services that handle heavy traffic with low resource usage.

Popular frameworks
GinEchoFiberChi
Rust
Performance · Systems

Extremely fast and memory-efficient. Steep learning curve. Used for critical systems and tools where every millisecond counts.

Popular frameworks
AxumActixTauri

Glossary

AI Glossary

Essential AI terminology explained in Swedish and English.

Filter by domain
Showing 243 of 243 terms

.env file

A local file containing secret values (API keys, database config). Never committed to Git. Used by code to read secrets without exposing them in source code.
SEE ALSO:
DOMAIN:

.gitignore

A Git configuration file that specifies which files should NOT be pushed to GitHub. .env files must be here to prevent secret keys from leaking.
SEE ALSO:
DOMAIN:

a11y

Numeronym for accessibility — building products usable by everyone, including people with disabilities. Screen readers, keyboard navigation, color contrast, focus indicators, ARIA attributes. WCAG 2.1 AA is the standard bar for EU services.
SEE ALSO:
DOMAIN:

Acceptance Criteria

The specific conditions that must be met for a task or feature to be considered done. Clear acceptance criteria make success measurable.
SEE ALSO:
DOMAIN:

Agent

An autonomous AI system that can plan, decide, and act to achieve a goal. Agents use tools, memory, and feedback loops to operate independently.
SEE ALSO:
DOMAIN:

Agentic AI

AI systems that take multiple steps to solve a problem but still ask you at major decision points. Cursor and Lovable copilots are agentic — they plan and build, but wait for approval.
SEE ALSO:
DOMAIN:

Agent Loop

The recurring cycle where the agent plans, acts, observes the result, and adjusts the next step. It's the core of many agent architectures.
SEE ALSO:
DOMAIN:

AI Agent

An AI system that doesn't just answer, but can plan, use tools, and execute steps toward a goal. An agent combines a model, instructions, tools, memory, and control flow.
SEE ALSO:
DOMAIN:

AI Workflow

A defined workflow where AI helps with one or more steps. A workflow is often easier to productize than a completely open-ended chatbot.
SEE ALSO:
DOMAIN:

Anonymization

Process of removing or replacing identifying information from data so it can no longer be linked to a person. Like changing "Anna, 34, Gothenburg" to "female user, 30s-40s, major city". Significantly reduces GDPR risk.
SEE ALSO:
DOMAIN:

Alignment

Getting the model to behave in line with human goals, rules, and safety requirements. Alignment is about technology, ethics, and product design alike.
SEE ALSO:
DOMAIN:

Rate Limit

A limit on how often a system can be called. Rate limits protect against overload, abuse, and unexpected costs.
SEE ALSO:
DOMAIN:

Anthropic

US-based AI provider behind Claude. Data stored in US and subject to CLOUD Act. Requires DPA for GDPR compliance.
SEE ALSO:
DOMAIN:

Anti-pattern

A solution that looks reasonable but creates more problems than it solves — the opposite of a design pattern. Retry storm is a clear example: the rule "retry on failure" seems right until the first system hiccup turns every client into a kamikaze against your own server.
SEE ALSO:
DOMAIN:

User Prompt

The instruction the user writes directly to the model. It is interpreted within the bounds of the system's higher-level instructions and safety rules.
SEE ALSO:
DOMAIN:

Use Case

A concrete use case, such as "build a research agent" or "automate support". Use cases make AI easier to understand than abstract concepts.
SEE ALSO:
DOMAIN:

API

Application Programming Interface — a way for one software to talk to another. An API lets you use another service's features without needing to build them yourself.
SEE ALSO:
DOMAIN:

API key

A secret string that proves you are authorized to use an API. Treat it like a password — don't share it, and rotate it regularly.
SEE ALSO:
DOMAIN:

Workflow

A chain of steps from input to result. In AI products a workflow can include model runs, tools, checks, and human approvals.
SEE ALSO:
DOMAIN:

Workflow Agent

An agent that works according to a defined flow — gather input, analyze, execute, verify, and report. It's often more robust than a fully open-ended agent.
SEE ALSO:
DOMAIN:

Assistant Design

The design of how an AI assistant should behave, respond, and help the user. It covers UX, tone, boundaries, and functionality alike.
SEE ALSO:
DOMAIN:

Source of Truth

The source the system should treat as authoritative. In AI products it matters to know whether the model should trust documentation, the database, or user input.
SEE ALSO:
DOMAIN:

Authentication

The process of verifying that someone (or a system) is who they claim to be. Usually done with passwords, keys, or other secrets.
SEE ALSO:
DOMAIN:

Automation

Letting a machine run a flow you would otherwise do manually. AI automation goes further: the machine interprets, decides, and adapts instead of just following fixed rules. Tools like n8n, Zapier and Make orchestrate automation across services — with or without LLMs in the loop.
SEE ALSO:
DOMAIN:

Autonomous Agent

An agent that can take multiple steps on its own without asking the user every time. Autonomy requires strong boundaries, logging, and a way to stop the agent.
SEE ALSO:
DOMAIN:

Autonomy

The ability of an AI system to make decisions and take action without human guidance at each step. Agents have autonomy; chatbots do not.
SEE ALSO:
DOMAIN:

Balanced model

Mid-tier models with good quality at reasonable cost. Claude Sonnet, GPT-4o mini, Gemini Pro. Most common in agent systems — solid results without wasting budget.
SEE ALSO:
DOMAIN:

Bash

Bourne Again Shell — the most common shell on Linux and WSL, the default on Mac until 2019 as well. When a tutorial says "run this in the terminal", it usually means bash or zsh. Also the language you write shell scripts in — automation that does not need Python.
SEE ALSO:
DOMAIN:

Rubric

A grading template describing what a good answer must contain. Rubrics make AI evaluation more consistent.
SEE ALSO:
DOMAIN:

Benchmark

A standardized test used to compare models or systems. Benchmarks are useful but don't always capture real product needs.
SEE ALSO:
DOMAIN:

Dependency

External package your code needs to work. Listed in `package.json` (JavaScript), `requirements.txt` (Python), or `Cargo.toml` (Rust). Each dependency pulls in its own dependencies — a single `npm install` can fetch hundreds of packages. Also a supply chain risk: a hacked dependency drags your entire code with it.
SEE ALSO:
DOMAIN:

Bucket

Storage container for files in the cloud — images, videos, documents, model weights, anything but traditional databases. AWS S3, Google Cloud Storage and Cloudflare R2 are the most common. Used by AI systems to store training data, models, and uploaded user content. Note jurisdiction: a bucket in the US is subject to US laws.
SEE ALSO:
DOMAIN:

Caching

Saving previous results or intermediate values to avoid recomputing the same thing. Caching can reduce both cost and latency.
SEE ALSO:
DOMAIN:

Chroma

A vector database commonly used for RAG. It's easy to set up locally, supports embedding models, and can work both in-memory and persistently on disk.
SEE ALSO:
DOMAIN:

Chunk

A piece of text from a larger document, cut down to fit into the RAG system. A good chunk is 400–800 tokens with 10–20% overlap with neighbors. Too large chunks dilute the signal, too small ones lose context.
SEE ALSO:
DOMAIN:

Chunking

Splitting documents into smaller pieces the model can use. Good chunking is critical for RAG to find the right information.
SEE ALSO:
DOMAIN:

CI/CD

Continuous Integration/Continuous Deployment — automation that tests and deploys your code automatically when you push to GitHub. Saves time and reduces bugs.
SEE ALSO:
DOMAIN:

CLI

Command Line Interface — text-based interface where you type commands instead of clicking. Faster than a GUI once you learn the basics, and necessary for many AI tools: Claude Code, Ollama, n8n, and Docker all run via CLI. Lives in a terminal app.
SEE ALSO:
DOMAIN:

CLOUD Act

Clarifying Lawful Overseas Use of Data Act — US law giving US authorities the right to request data from US companies, even if stored in Europe. Affects OpenAI, Anthropic, Google, and other US-based services.
SEE ALSO:
DOMAIN:

Cloudflare

Edge platform with CDN, DNS, R2 storage, and Workers AI. Can run AI models in 300+ data centers close to users. The free tier is generous and fits small-scale AI projects.
SEE ALSO:
DOMAIN:

Container

An isolated environment containing everything an application needs to run. Like a room in a house — shares walls and foundation with the rest, but can be easily shut down without affecting other rooms.
SEE ALSO:
DOMAIN:

Data Quality

How good, correct, complete, and relevant the data is. Poor data often produces poor AI results, even when the model is strong.
SEE ALSO:
DOMAIN:

Data Retention

How long data is stored and when it is removed. This matters in AI systems because prompts, files, and logs can contain sensitive information.
SEE ALSO:
DOMAIN:

Dataset

A structured collection of data used for training, evaluation, or analysis. For AI systems, the dataset's quality often matters more than its size.
SEE ALSO:
DOMAIN:

Data sovereignty

The principle that data is subject to the legal jurisdiction where the server is located. EU storage = EU laws, US storage = US laws. Critical for compliance, privacy, and legal risk.
SEE ALSO:
DOMAIN:

Data Exfiltration

When data leaks out of a system to the wrong recipient. In AI systems this can happen through prompts, tools, faulty access, or poor isolation.
SEE ALSO:
DOMAIN:

DeepSeek

Chinese AI provider that has released open-weight reasoning models (R1, V3) that match GPT-4 at a fraction of the cost. The models are free to download — but the servers are in China.
SEE ALSO:
DOMAIN:

Docker

Container technology that packages your code, libraries, and settings in a box that works the same everywhere — on laptop, server, or cloud. Standard for deploying AI agents and software.
SEE ALSO:
DOMAIN:

DPA

Data Processing Agreement — a legal contract between you and an AI provider that confirms how personal data is handled. Required for GDPR compliance when data is sent to the US.
SEE ALSO:
DOMAIN:

Edge

Running code in data centers close to the user rather than one central location. Cloudflare has 300+ edge locations worldwide; Vercel and Netlify build on the same principle. For AI: edge inference lets the model respond with low latency wherever the user is — useful for chat UIs and real-time agents.
SEE ALSO:
DOMAIN:

Embedding

A way to turn text into numbers that a model can understand. Words with similar meanings get similar numbers, so the model can understand what things are related.
SEE ALSO:
DOMAIN:

Embedding model

An AI model that converts text to vectors. multilingual-e5-large is a popular embedding model that can handle 100+ languages. Many agents use it to search through their own memory.
SEE ALSO:
DOMAIN:

Escalation

When the AI system hands over to a human or another system. It's needed when the question is too sensitive, uncertain, or outside the agent's authority.
SEE ALSO:
DOMAIN:

Example-based Steering

Steering the model by showing good and bad examples. This is often stronger than describing abstract rules alone.
SEE ALSO:
DOMAIN:

fastembed

A fast embedding library built in Rust. Runs locally without needing a separate embedding API. Popular in the MCP layer to hide embedding details from the agent.
SEE ALSO:
DOMAIN:

Few-shot Prompting

When you give the model a few examples of what the answer should look like. It helps the model understand the pattern, tone, structure, and quality bar.
SEE ALSO:
DOMAIN:

Fine-tuning

Continuing to train a model on specific data to change its behavior or specialize it. Fine-tuning is powerful, but isn't always needed — sometimes better prompting or RAG is enough.
SEE ALSO:
DOMAIN:

Format Validation

Checking that the model's response follows the right structure. This can mean validating JSON, headings, fields, or length.
SEE ALSO:
DOMAIN:

Research Agent

An agent that searches for information, compares sources, and assembles conclusions. It especially needs good source handling so as not to mix facts, assumptions, and guesses.
SEE ALSO:
DOMAIN:

Frontier model

The most powerful, expensive AI models available — GPT-4o, Claude Opus, Gemini Ultra. Best for complex reasoning and difficult problems but slow and expensive to run at scale.
SEE ALSO:
DOMAIN:

Function Calling

A structured way for the model to call functions with defined arguments. It's used to connect LLMs to real systems and APIs.
SEE ALSO:
DOMAIN:

Skill Tree

A visual or logical map of skills the user can learn. It fits AI education well, since the field can otherwise feel overwhelming to navigate.
SEE ALSO:
DOMAIN:

Grounding

Anchoring the model's answer in concrete data, sources, or tool results. Grounding reduces the risk of hallucinations.
SEE ALSO:
DOMAIN:

GDPR

General Data Protection Regulation — EU law regulating how personal data is handled. Violations can result in fines up to 4% of global revenue.
SEE ALSO:
DOMAIN:

Generative AI

AI systems that can create new content, such as text, images, code, audio, or video. Unlike older AI systems that only classify or search, generative AI produces something new based on patterns it has learned.
SEE ALSO:
DOMAIN:

Git

Version control system that tracks changes to code over time. Created by Linus Torvalds. Lets you go back, compare versions, and collaborate without overwriting others. The foundation of modern development flows — and what enables AI agents like Claude Code to suggest and review changes safely.
SEE ALSO:
DOMAIN:

GitHub

Platform for storing and collaborating on code via Git. Owned by Microsoft. Most open source projects — and AI tools like Claude Code, OpenAI SDK, and n8n — are distributed through GitHub. Also where you publish your code to deploy to platforms like Vercel or Railway.
SEE ALSO:
DOMAIN:

Approval Gate

A checkpoint where the agent must wait for approval. Examples: "may I send the email?", "may I delete the file?", or "may I run the deploy?".
SEE ALSO:
DOMAIN:

Google

Tech giant that owns the Gemini model family, Google AI Studio, and Vertex AI. Deeply integrated with Workspace (Gmail, Docs, Sheets). Data centers globally — note this for GDPR-sensitive data.
SEE ALSO:
DOMAIN:

GPU

Graphics Processing Unit — hardware specialized in parallel computation. AI models run much faster on GPU than CPU. Nvidia GTX/RTX are most popular for local work.
SEE ALSO:
DOMAIN:

Critic

An agent role that reviews a proposal and looks for weaknesses, errors, or risks. A critic is often used to raise quality before something is published or run.
SEE ALSO:
DOMAIN:

Hallucination

When an AI model confidently gives an answer that sounds right but is actually wrong or made up. The model doesn't know it's wrong — it just generated something plausible.
SEE ALSO:
DOMAIN:

Secret management

Practice of securely storing and distributing secret values (keys, passwords, tokens). Tools: HashiCorp Vault, 1Password, AWS Secrets Manager. Never in plain text.
SEE ALSO:
DOMAIN:

Hugging Face

Platform for open source AI. Hosts thousands of models, datasets, and spaces (live demos). The de facto place to find and share AI models. Can be run locally or through their inference API.
SEE ALSO:
DOMAIN:

Retrieval

The process where the system fetches relevant information for the model. Retrieval can come from documents, databases, web search, or internal systems.
SEE ALSO:
DOMAIN:

Event

A signal that something has happened, such as "new customer created" or "file uploaded". Events are used to trigger workflows and agents.
SEE ALSO:
DOMAIN:

Event-driven

Architecture where components react to events rather than running on a fixed schedule. When an email arrives, when a file is uploaded, when a user clicks — a flow is triggered. n8n, Zapier, and AWS Lambda are typical event-driven systems. Suits AI agents meant to react to the world, not poll it.
SEE ALSO:
DOMAIN:

i18n

Numeronym for internationalization — preparing an app so it CAN be displayed in multiple languages. Infrastructure work: locale routing, message files, translation hooks. Distinct from l10n, which is the actual translation work. Agentakademin uses next-intl with /sv/ and /en/ routing.
SEE ALSO:
DOMAIN:

IaaS

Infrastructure as a Service — raw servers, storage, and networking in the cloud that you configure and operate yourself. AWS EC2, Google Compute Engine, and Hetzner Cloud are IaaS. Maximum control and flexibility, but you are responsible for the OS, security, and updates.
SEE ALSO:
DOMAIN:

Content Path

A recommended path through content, such as "start here, continue here, then build this". It saves the user from having to choose among too many cards.
SEE ALSO:
DOMAIN:

Instruction Hierarchy

The priority order between different instructions — system instructions, developer instructions, and the user's prompt. It keeps the model from simply following the most recent text it sees.
SEE ALSO:
DOMAIN:

Integration

The connection between the AI system and an external system, such as Slack, Gmail, GitHub, or a database. Integrations are often where AI goes from demo to real product.
SEE ALSO:
DOMAIN:

Privacy

Protection of users' personal data and sensitive information. AI systems should minimize unnecessary data collection and be transparent about what is stored.
SEE ALSO:
DOMAIN:

Interactive Hero

An interactive first section on a website that lets the user do something right away. For an AI learning site this could be a scan, diagnosis, or guide to the right content.
SEE ALSO:
DOMAIN:

Onboarding

The first experience where the user learns what the AI product does and how to use it. For AI products onboarding matters extra because the user often doesn't know what's possible.
SEE ALSO:
DOMAIN:

Isolation

Keeping a process separated from the rest of the system so that if something goes wrong — or if the code is malicious — the rest is not affected. Containers, virtual machines, and sandboxes are all forms of isolation. For AI agents, isolation is critical: an agent allowed to run code should never reach your secrets or production data.
SEE ALSO:
DOMAIN:

Iteration

Improving something step by step based on tests, feedback, or new insights. AI products should almost always be built iteratively because behavior can be hard to predict.
SEE ALSO:
DOMAIN:

Jailbreak

An attempt to make the model break its rules. Jailbreaks can happen through manipulation, role-play, or instructions that try to bypass safety systems.
SEE ALSO:
DOMAIN:

JavaScript

The language of the web. Runs in the browser and on Node.js servers. You can't build a web UI without running into JavaScript — but for AI logic, Python and TypeScript are more common.
SEE ALSO:
DOMAIN:

JSON Mode

A mode where the model is expected to respond with valid JSON. It's often used when AI output is going to be passed on to program code.
SEE ALSO:
DOMAIN:

k8s

Numeronym for Kubernetes — the open-source system for orchestrating containers at scale. The 8 stands for the letters between k and s ("ubernete"). Mostly used in devops contexts: "we run k8s in prod". Overkill for a one-person project, invaluable for a 100-person team with 50 microservices.
SEE ALSO:
DOMAIN:

Calibration

How well the model's confidence matches actual correctness. A calibrated model is cautious when it should be cautious.
SEE ALSO:
DOMAIN:

Coding Agent

An agent that can read, write, change, and test code. Good coding agents need clear goals, repo context, tests, and protection against incorrect file changes.
SEE ALSO:
DOMAIN:

Compression

Reducing the number of tokens by summarizing, paraphrasing, or picking the most important parts. Compression lets more information fit inside the token budget.
SEE ALSO:
DOMAIN:

Confabulation

A more technical word for when the model fills gaps with something that sounds plausible but is unverified. It can be dangerous because the answer often sounds convincing.
SEE ALSO:
DOMAIN:

Confidence

An estimate of how certain the model or system is about an answer. Confidence should not be confused with actual truth — a self-assured answer can still be wrong.
SEE ALSO:
DOMAIN:

Context

Information the model needs to understand the question or task. Good context lets the model give more relevant and accurate answers.
SEE ALSO:
DOMAIN:

Context window

The amount of information the model can actively have in front of itself at once. A larger context window allows the model to read longer documents, more code, or longer conversations.
SEE ALSO:
DOMAIN:

Cost

The cost of running the model, often based on number of tokens, model choice, and tool usage. Cost becomes important fast once AI products gain many users.
SEE ALSO:
DOMAIN:

Encryption

Scrambling data so that only someone with the right key can read it. Like putting a message in a locked box that only you have the key to.
SEE ALSO:
DOMAIN:

Kubernetes

Orchestration system for containers. Manages many containers, updates, scaling, and automatic restart. Used when you need to run many agents or handle high traffic.
SEE ALSO:
DOMAIN:

Knowledge Base

A collected knowledge source the AI system can search. It can be documentation, support articles, internal rules, or product information.
SEE ALSO:
DOMAIN:

Course Module

A standalone unit within a course or learning site. Course modules can be themed, level-based, or tool-specific.
SEE ALSO:
DOMAIN:

Quantization

The process of compressing an AI model by reducing the precision of the numbers inside it. A model is quantized to Q4 (4-bit), Q8 (8-bit), or other levels to fit smaller GPU memory. Lower precision = smaller file and faster execution, but marginally lower quality.
SEE ALSO:
DOMAIN:

Citation

A source reference showing where the information comes from. Citations matter when the user needs to verify the answer.
SEE ALSO:
DOMAIN:

l10n

Numeronym for localization — adapting content to a specific language or region. Translating strings, but also date formats, currency, formal/informal register, and idioms that don’t translate directly. The l10n work for Agentakademin was keeping the Swedish tone in English without losing nuance.
SEE ALSO:
DOMAIN:

Latency

The time it takes from when the user sends a query until the answer starts arriving or is complete. Low latency matters for AI products to feel fast.
SEE ALSO:
DOMAIN:

Linux

Open operating system that is standard on servers and in the cloud. Standard for agents, Docker, and all development tools. Bash scripting and command-line become your daily language.
SEE ALSO:
DOMAIN:

LLM

A large language model trained on enormous amounts of text and code to understand and generate language. LLMs power tools like ChatGPT, Claude, and many AI agents.
SEE ALSO:
DOMAIN:

LLMjacking

An attack where adversaries steal an LLM API key and use it to run requests at the victim's expense. They don't steal data, they steal usage. A leaked Gemini key reportedly turned a ~$180/month bill into $82,314 in 48 hours. The risk grows as businesses embed language models into websites, apps, and internal tools: if a key leaks, the attacker doesn't need to access data, they just run usage on the account until the bill explodes.
SEE ALSO:
DOMAIN:

LM Studio

Desktop app for running LLMs locally on your own computer. Offers a built-in model library, chat UI, and OpenAI-compatible API. Easier for beginners than Ollama, but limited to a GUI.
SEE ALSO:
DOMAIN:

Logging

Saving information about what the system did — prompts, tool calls, errors, and results. Logging is necessary for debugging, quality, and security.
SEE ALSO:
DOMAIN:

Locally

Running code and models on your own computer or server instead of in the cloud. Advantages: no data leaves your machine, GDPR-compliant, no cloud cost. Disadvantages: limited hardware, you must update yourself.
SEE ALSO:
DOMAIN:

Lossy AI

An AI model that deliberately loses information to reduce size, speed, or costs. Common when quantizing models (Q4, Q8) or trimming the context window. The opposite of lossless, where no information is lost.
SEE ALSO:
DOMAIN:

Learning Path

A structured learning path from foundational to advanced level. For AI and agents it often runs from concepts to practical workflows and then robustness.
SEE ALSO:
DOMAIN:

Make

Visual automation platform (formerly Integromat) that lets you build complex flows with AI models, data transformation, and integrations. Clearer visual model than Zapier — fits you if you want to see the entire chain.
SEE ALSO:
DOMAIN:

MCP

Model Context Protocol — standard for connecting agents to external tools and data sources. Makes it easy for an agent to read, write, and call APIs without you writing integration code yourself.
SEE ALSO:
DOMAIN:

Meta

The company behind the Llama model, PyTorch, and Facebook/Instagram. Llama is open-weight — you can download it and run it yourself. A larger contributor to AI research than many realize.
SEE ALSO:
DOMAIN:

Microsoft

Tech giant behind Copilot, Azure OpenAI Service, and Playwright. Has invested heavily in OpenAI and has its own AI stack (Phi models, Bing Chat). EU data centers are available for GDPR-sensitive workloads.
SEE ALSO:
DOMAIN:

Abuse Prevention

Protection against abuse, spam, overload, or harmful usage patterns. AI products need this because models can be expensive and powerful.
SEE ALSO:
DOMAIN:

Mistral

French AI provider with servers in the EU. GDPR-friendly alternative to OpenAI and Anthropic. Can be run locally via Ollama or via cloud API.
SEE ALSO:
DOMAIN:

Model

The AI system itself that takes input and produces output. A model can be specialized in text, images, audio, code, or multiple formats at once.
SEE ALSO:
DOMAIN:

Model Drift

When the model's behavior or quality changes over time. Drift can come from a new model version, changed data, changed prompts, or changing user needs.
SEE ALSO:
DOMAIN:

Model Routing

Choosing different models for different tasks. A cheaper model can handle simple steps while a stronger model is used for harder reasoning.
SEE ALSO:
DOMAIN:

Module

A standalone component or unit. In AI, modules can refer to different parts of an agent, model, or learning site.
SEE ALSO:
DOMAIN:

Multi-agent system

Multiple AI agents collaborating to solve a task too complex for one agent alone. One agent writes code, another reviews, a third tests. CrewAI, AutoGen, and LangGraph are frameworks for building multi-agent systems. Requires good orchestration — otherwise the agents start talking over each other.
SEE ALSO:
DOMAIN:

multilingual-e5-large

An embedding model that handles 100+ languages and produces 1024-dimensional vectors. Popular for Swedish agents and RAG systems that need multilingual support without model switching.
SEE ALSO:
DOMAIN:

Multimodal Model

A model that can handle multiple types of input, such as text, images, audio, or video. This lets the model understand more of a situation than just written text.
SEE ALSO:
DOMAIN:

Minimum Viable Product (MVP)

Minimum Viable Product: the smallest version that can deliver real value and be tested with users. In AI products the MVP should prove the usefulness, not just show off the model.
SEE ALSO:
DOMAIN:

Goal State

The desired end state of a task. The clearer the goal state, the easier it is for the agent to know when it's done.
SEE ALSO:
DOMAIN:

Human-in-the-Loop

When a human must review or approve a step before the AI system continues. It's often used for risky decisions or external actions.
SEE ALSO:
DOMAIN:

Human Feedback

Human judgment of the model's responses or behavior. Feedback can be used for training, evaluation, or product improvement.
SEE ALSO:
DOMAIN:

n8n

Open source platform for automation and AI workflows. You build flows visually — connecting tools, APIs, agents, and triggers without writing code. Can be run locally on your own server for full data control.
SEE ALSO:
DOMAIN:

npm

Node Package Manager — the largest registry for JavaScript packages, with over 2 million packages. Owned by GitHub (Microsoft). When you run `npm install <package>` code is fetched from npm servers. Also a supply chain risk: a compromised package can inject code into thousands of applications in hours.
SEE ALSO:
DOMAIN:

npx

Command that runs npm packages without installing them permanently. `npx <package>` fetches, runs, and cleans up. Handy for one-off tools or trying something: `npx create-next-app` creates a new Next.js project without you having to install the CLI globally.
SEE ALSO:
DOMAIN:

Numeronym

An abbreviation where the digit in the middle counts the letters stripped out. i18n means internationalization (18 letters between i and n). Common in tech: i18n (internationalization), l10n (localization), a11y (accessibility), k8s (kubernetes), m17n (multilingualization). Useful when words are long and used often — and esoteric enough to signal insider status.
SEE ALSO:
DOMAIN:

Observation

Information the agent gets back after an action — a tool's result, an error message, or a search hit. Observations drive the next step in the agent loop.
SEE ALSO:
DOMAIN:

Ollama

Tool for running local AI models (Llama, Mistral, Qwen) on your own computer or server. No data leaves the machine — right choice for sensitive data and GDPR compliance.
SEE ALSO:
DOMAIN:

Reranking

When search results are reordered by how relevant they actually are to the query. Reranking is often used to raise quality in RAG systems.
SEE ALSO:
DOMAIN:

OpenAI

US-based AI provider behind ChatGPT. Data stored in US and subject to CLOUD Act. Requires DPA for GDPR compliance.
SEE ALSO:
DOMAIN:

Operating system

The program that controls a computer and lets other programs run on it. Linux (Ubuntu, Debian), macOS, and Windows are the three big ones. For AI development most run Linux or WSL (Linux on Windows) — that is where all tools are written to run. macOS works well too. Pure Windows is still awkward for some AI tools.
SEE ALSO:
DOMAIN:

Orchestration

Coordinating multiple components — agents, tools, models — so they work together toward a goal. An orchestrator decides who does what and in what order. LangGraph, CrewAI, and n8n are orchestration tools. The difference between a demo and a production AI system often lies in how well the orchestration is done.
SEE ALSO:
DOMAIN:

Uncertainty

Uncertainty in the model's answer. Good AI systems should be able to signal when they don't know, when they're guessing, and when something needs to be verified.
SEE ALSO:
DOMAIN:

PaaS

Platform as a Service — a platform where you push your code and the platform handles servers, deployment, scaling, and updates. You focus on the code, not the infrastructure. Vercel, Railway, Fly.io, and Heroku are classic PaaS. More control than SaaS, less work than IaaS.
SEE ALSO:
DOMAIN:

Persona

The style, role, or identity given to an AI system in a particular product. A persona can shape tone and priorities, but shouldn't replace actual functionality.
SEE ALSO:
DOMAIN:

Personalization

When the system adapts content or recommendations to the user's goal, level, or behavior. It makes AI learning more relevant than a static course catalog.
SEE ALSO:
DOMAIN:

Personal data

Any information that can be linked to an identifiable person: name, email, phone number, IP address, cookies, ID number. GDPR-protected and requires extra care when sent to AI models.
SEE ALSO:
DOMAIN:

pip

Package Installer for Python — the command you run to install Python packages from the PyPI registry. `pip install openai` adds the OpenAI library to your environment. For AI development in Python, pip is what npm is for JavaScript. Use virtual environments (`venv`) to keep projects separate.
SEE ALSO:
DOMAIN:

Planner

The part of the agent that breaks a goal into steps. A good planner makes the task more manageable and reduces the risk of the agent jumping straight to the wrong solution.
SEE ALSO:
DOMAIN:

Playwright

Microsoft tool for controlling browsers programmatically. AI agents use Playwright to navigate pages, click buttons, and extract data — like a human but a hundred times faster. Often exposed as an MCP server.
SEE ALSO:
DOMAIN:

Podman

Open source alternative to Docker. Same commands, same container format, but daemon-less and rootless by default — safer for production. Drop-in replacement for most Docker flows.
SEE ALSO:
DOMAIN:

Policy

A set of rules for how the AI system should behave. Policies can govern safety, tone, tool usage, data protection, and escalation.
SEE ALSO:
DOMAIN:

Precision

How much of the search results are actually relevant. High precision means the user gets less noise.
SEE ALSO:
DOMAIN:

Prompt

The instruction or question sent to the model. A good prompt describes the goal, context, constraints, and desired format.
SEE ALSO:
DOMAIN:

Prompt engineering

The work of formulating instructions so the model gives better, more useful, and more consistent answers. It's less about magic phrases and more about clarity, context, and control.
SEE ALSO:
DOMAIN:

Prompt Injection

When malicious or unexpected text tries to make the model ignore its instructions. It's especially dangerous in RAG systems and agents that read external documents.
SEE ALSO:
DOMAIN:

Prompt Template

A reusable template for how prompts should be built. Templates are common in products where the model is doing the same kind of task many times.
SEE ALSO:
DOMAIN:

Prototype

An early version that shows how something could work. Prototypes are used to test ideas quickly before building the finished product.
SEE ALSO:
DOMAIN:

Python

Programming language that dominates AI and data. Most AI libraries (PyTorch, Transformers, LangChain) are written in Python. Readable syntax — a good first language if you want to build with AI in depth.
SEE ALSO:
DOMAIN:

Q4

A 4-bit quantization — the most compressed variant. Gives the smallest file size and fits most GPUs. Quality loss is marginal for most tasks.
SEE ALSO:
DOMAIN:

Q8

An 8-bit quantization — better than Q4 but larger file. About twice as large as Q4, but closer to full quality. A good middle ground between size and precision.
SEE ALSO:
DOMAIN:

Qdrant

A high-performance vector database built in Rust. Better at scaling and complex queries than Chroma. Popular when you need a production-grade RAG system.
SEE ALSO:
DOMAIN:

RAG

Retrieval Augmented Generation — a technique where the model looks up relevant documents or data before answering a question. This helps it give more accurate answers without retraining.
SEE ALSO:
DOMAIN:

Railway

Deployment platform for agents, APIs, and databases with one-click deploy from GitHub. The free tier is enough for MVPs. Runs in the US — note this for GDPR-sensitive data.
SEE ALSO:
DOMAIN:

Reflection

When the agent analyzes its own work and tries to improve the next step. Reflection can help, but doesn't replace tests or hard verification.
SEE ALSO:
DOMAIN:

Compliance

An AI system following laws, rules, and internal requirements. For companies this can cover GDPR, security, logging, data retention, and accountability.
SEE ALSO:
DOMAIN:

Regression Test

A test that checks something that worked before still works. It matters because AI systems can easily degrade when prompts, models, or data change.
SEE ALSO:
DOMAIN:

Recommendation

A suggested next action, course, article, or path based on the user's input. Good recommendations are clear, justified, and not overloaded.
SEE ALSO:
DOMAIN:

Recursive text splitting

Standard chunking method — splits at paragraphs → sentences → words until chunk size is reached. Preserves semantic coherence and is implemented in LangChain and LlamaIndex.
SEE ALSO:
DOMAIN:

Fallback

A backup path when something goes wrong, such as switching models, asking the user to simplify, or sending the case to a human. Fallbacks make AI systems more robust.
SEE ALSO:
DOMAIN:

Reasoning model

An AI model that takes time to think through complex problems before answering. It writes internal notes, checks itself, and approaches the problem from multiple angles. Useful for mathematics, logic, and difficult analysis.
SEE ALSO:
DOMAIN:

Audit Log

A log that shows who or what performed a given action and when. Audit logs matter when AI systems gain access to sensitive functions.
SEE ALSO:
DOMAIN:

Reinforcement Learning from Human Feedback (RLHF)

Reinforcement Learning from Human Feedback: a method where human feedback is used to train the model to give better answers. It's often used to make models more helpful and safer.
SEE ALSO:
DOMAIN:

Role Prompting

When the model is given a role, such as "act as a lawyer", "act as a code reviewer", or "act as a teacher". The role can help, but must be combined with clear goals and rules.
SEE ALSO:
DOMAIN:

Fairness

How fairly an AI system treats different users, groups, or cases. Fairness requires both technical evaluation and clear value judgments.
SEE ALSO:
DOMAIN:

S3

Amazon's "Simple Storage Service" — the original cloud object storage, launched 2006. Standard place to store files for AI applications: training data, model weights, generated images. "S3-compatible API" has become a de facto standard implemented by Cloudflare R2, MinIO and others.
SEE ALSO:
DOMAIN:

SaaP

Software as a Product — software you buy or download and own as a product, rather than renting it as a service. Think classic desktop programs, or AI models you download and run locally with Ollama or LM Studio. The opposite of the SaaS subscription model.
SEE ALSO:
DOMAIN:

SaaS

Software as a Service — ready-made software you use in a browser or app, without installing or operating anything. You pay monthly or annual subscriptions. ChatGPT, Notion, Slack, and Spotify are SaaS. Fastest to start, but you own neither the data nor the infrastructure.
SEE ALSO:
DOMAIN:

Sandboxing

Technique of isolating an agent or code so it can only access what it needs — no access to the whole system. Like letting someone work in a small room instead of the entire office.
SEE ALSO:
DOMAIN:

Sandbox

An isolated environment where the agent can work without harming real systems. Sandboxes are often used for code execution, file changes, and testing.
SEE ALSO:
DOMAIN:

Schema Validation

Checking that input or output follows a defined schema. It matters when AI systems are meant to be reliable in production.
SEE ALSO:
DOMAIN:

SearXNG

EU-friendly metasearch engine that aggregates results from Google, Bing, and others without storing user data. Can be self-hosted — fits agents that need real-time search without leaking queries to third parties.
SEE ALSO:
DOMAIN:

Semantic chunking

Chunking based on embedding distance instead of character count. Creates new chunks when the subject "changes topic". Gives higher precision but requires an extra embedding pass during indexing.
SEE ALSO:
DOMAIN:

Shell

The program inside the terminal that interprets and runs your commands. The most common: bash (Linux default), zsh (Mac default since 2019), fish (modern, more user-friendly). When you type `ls`, the shell is what lists files for you.
SEE ALSO:
DOMAIN:

Self-correction

When the agent detects an error and tries to fix it on its own. It works best when the agent has clear signals — failed tests or validation rules, for example.
SEE ALSO:
DOMAIN:

Scaling

Handling more load without the system crashing. Vertical scaling = bigger server. Horizontal scaling = more servers sharing the load. Modern PaaS like Vercel and Railway scale automatically: as traffic rises, more instances spin up. For AI systems scaling matters extra because each LLM call is expensive — if your app goes viral, costs can spiral.
SEE ALSO:
DOMAIN:

Guardrail

A rule, block, or check that limits what the AI system is allowed to do. Guardrails can be technical, legal, or product-level.
SEE ALSO:
DOMAIN:

Service Level Agreement (SLA)

Service Level Agreement: a promise of availability or performance. In AI products, SLAs can be tricky because external models and tools are often part of the picture.
SEE ALSO:
DOMAIN:

Fast model

Small, cheap models for simple tasks at high volume. Claude Haiku, Gemini Flash, GPT-3.5. Best for repetitive work, classification, and quick iteration.
SEE ALSO:
DOMAIN:

Bias

Systematic skew in data, model, or output. Bias can cause AI systems to treat groups, topics, or situations incorrectly.
SEE ALSO:
DOMAIN:

Trace

A detailed run history showing which steps an AI system took. Traces make it easier to understand why an agent went wrong.
SEE ALSO:
DOMAIN:

Structured Output

When the model is forced or steered to respond in a defined structure. It matters when AI responses are going to be used by code, APIs, or other systems.
SEE ALSO:
DOMAIN:

Streaming

When the response appears incrementally while the model is generating it. Streaming makes the product feel faster even if the full answer takes time.
SEE ALSO:
DOMAIN:

sudo

"Superuser do" — the command you put in front of another when it requires admin privileges on Linux or Mac. `sudo apt install <package>` installs system-wide. Use carefully: sudo grants full access to the entire computer. Never copy-paste `sudo` commands without understanding what they do.
SEE ALSO:
DOMAIN:

Supply chain

The chain of dependencies and packages your code relies on. Risk: malicious code sneaked into a popular package. Defense: version pinning, scanning, and private registries.
SEE ALSO:
DOMAIN:

Support Agent

An agent that helps users or customers with questions, troubleshooting, or cases. It usually needs a knowledge base, escalation paths, and clear rules for what it must not do.
SEE ALSO:
DOMAIN:

Blacklisting

Trying to block forbidden things while allowing everything else. It can work in simple cases, but is weaker than whitelisting when the stakes are high.
SEE ALSO:
DOMAIN:

Vision Model

A model specialized in understanding and analyzing images. Vision models can identify objects, read text in images, or answer questions about image content.
SEE ALSO:
DOMAIN:

Synthetic Data

Data created artificially, often by a model or simulator. It can be useful, but risks amplifying errors if not checked carefully.
SEE ALSO:
DOMAIN:

System Prompt

A higher-level instruction that steers the model's behavior throughout the entire conversation or run. It's often used to define role, rules, safety, and response style.
SEE ALSO:
DOMAIN:

Security

In AI contexts: protection against your model, data, or keys being abused. Four pillars: API keys (rotation and scope), .env files (gitignore and vault), sandboxing (isolate the agent), and supply chain (audit dependencies). Security decides whether the rest of the stack holds up over time.
SEE ALSO:
DOMAIN:

Terminal

The app where you run commands — Terminal.app on Mac, Windows Terminal on Windows, GNOME Terminal or Alacritty on Linux. Inside the terminal runs a shell (usually bash or zsh). This is where you install and run most AI development tools.
SEE ALSO:
DOMAIN:

Test Set

A collection of examples used to test an AI system. A good test set resembles the problems the system is actually meant to solve.
SEE ALSO:
DOMAIN:

Test Runner

A tool or process that automatically runs tests and reports the results. Test runners are important for ensuring code or system quality.
SEE ALSO:
DOMAIN:

Reliability

How stably and reliably the system works over time. For AI, reliability is about both technical uptime and answers staying high quality.
SEE ALSO:
DOMAIN:

TLS

Transport Layer Security — the encryption protocol that secures almost all internet traffic today. It is what the "S" in HTTPS stands for. When you send an API key to OpenAI, TLS protects it from being read by others on the way. The modern standard is TLS 1.3. Without TLS, everything is sent in plain text.
SEE ALSO:
DOMAIN:

Token

A small piece of text that the model reads or writes. A token can be an entire word, part of a word, or punctuation.
SEE ALSO:
DOMAIN:

Token Budget

The amount of tokens that fits within a run, including both input and output. When the budget runs out, the model has to prioritize, summarize, or drop information.
SEE ALSO:
DOMAIN:

Tool calling

When an AI model decides to use a tool (like making an API call, reading a file, or sending an email) to solve a problem. The model itself decides when and how to use the tool.
SEE ALSO:
DOMAIN:

Accuracy

How often a model gives the right answer. Most often measured as percent correct on a test set. High accuracy on a benchmark does not mean the model is good at your task — context decides. For AI agents, consistent accuracy over time matters more than a top score on a single run.
SEE ALSO:
DOMAIN:

Training Data

The data used to train a model. The quality of the training data shapes what the model knows, how it behaves, and which biases it picks up.
SEE ALSO:
DOMAIN:

TypeScript

JavaScript with static type checking. Standard language for modern web apps and Next.js projects. Worth learning after Python if you want to build AI interfaces or MCP servers.
SEE ALSO:
DOMAIN:

Task Decomposition

Breaking a larger task into smaller, clearer steps. It's one of the most important techniques for getting agents to work stably.
SEE ALSO:
DOMAIN:

Output Format

The format the model should respond in, such as JSON, Markdown, a table, or a bulleted list. A clear format reduces the risk of unusable answers.
SEE ALSO:
DOMAIN:

Executor

The part of the agent that actually carries out the steps, such as calling tools, writing files, or making API requests. The executor should have clearly scoped permissions.
SEE ALSO:
DOMAIN:

Trigger

What starts an AI flow — a form submission, a button, a schedule, or a new email. Good triggers make automations more predictable.
SEE ALSO:
DOMAIN:

Evaluation

The process of measuring how well an AI system performs. It can cover correctness, helpfulness, safety, speed, or cost.
SEE ALSO:
DOMAIN:

Variables

Placeholders in a prompt or workflow, such as {customer-name}, {goal}, or {document}. Variables let the same prompt be used dynamically with different data.
SEE ALSO:
DOMAIN:

Vector

A list of numbers that represents something (a word, sentence, image) in mathematical space. Similar things have similar numbers — that is how semantic search works.
SEE ALSO:
DOMAIN:

Vector Database

A database built to store and search through embeddings. It's often used in RAG systems to find relevant context quickly.
SEE ALSO:
DOMAIN:

Vercel

Deployment platform optimized for Next.js, React, and edge functions. Push to Git and Vercel deploys automatically. Includes the AI SDK for running LLM calls on the server's edge nodes with low latency.
SEE ALSO:
DOMAIN:

Verifier

A component or role that checks whether the result is correct. In code it can be tests; in research, source checking; in workflows, rule validation.
SEE ALSO:
DOMAIN:

Verification

The process of verifying that something is correct or that a condition is met. Verification can mean testing code, validating data, or reviewing results.
SEE ALSO:
DOMAIN:

Tool Call

When the model asks the system to use a tool, such as searching the web, reading a file, or running code. Tool calls make the model action-capable but require oversight.
SEE ALSO:
DOMAIN:

Tool-using Agent

An agent that can use external tools — search, calendar, database, code environment, or API. Tool use makes the agent more capable but also riskier.
SEE ALSO:
DOMAIN:

Tool Permission

Rules for which tools an agent may use and under what conditions. Example: reading files is allowed, but deleting files requires approval.
SEE ALSO:
DOMAIN:

Tool Schema

The definition of how a tool may be used — which fields are required and which values are allowed. A good schema reduces the risk of incorrect calls.
SEE ALSO:
DOMAIN:

Tool Whitelist

A list of tools the agent is allowed to use. Whitelisting is often safer than trying to block everything dangerous after the fact.
SEE ALSO:
DOMAIN:

Whitelisting

Only allowing explicitly approved things — tools, domains, or commands. It's the opposite of blacklisting and is often safer.
SEE ALSO:
DOMAIN:

VPS

Virtual Private Server — a whole server just for you, rented from a hosting provider. From 99 SEK/month you can run an agent 24/7.
SEE ALSO:
DOMAIN:

VRAM

Video RAM — the memory a graphics card uses. Limited on most GPUs (typically 8-24 GB for consumer GPUs). Larger models and lower quantization require more VRAM.
SEE ALSO:
DOMAIN:

Webhook

A way for one system to automatically notify another system when something happens. Like a doorbell that calls your phone when someone rings.
SEE ALSO:
DOMAIN:

WSL

Windows Subsystem for Linux — Microsoft's official way to run Linux inside Windows. You get a full Linux system (usually Ubuntu) that shares files with Windows but runs Linux programs natively. For AI development on Windows, WSL is essentially mandatory — most AI tools are written for Linux and work poorly on pure Windows.
SEE ALSO:
DOMAIN:

WSL2

Windows Subsystem for Linux 2 — a real Linux kernel inside Windows. Lets you run Linux tools (Docker, Ollama, Python) directly from Windows without dual-boot.
SEE ALSO:
DOMAIN:

Zapier

The world's largest automation platform, with 6,000+ integrations. You connect AI to Gmail, Slack, CRM, and thousands of other tools without code. Closes the gap between AI models and existing work tools.
SEE ALSO:
DOMAIN:

Zero-shot Prompting

When the model gets a task without examples. It works well for simple or clear tasks, but weakens when the task requires a specific format or particular judgment.
SEE ALSO:
DOMAIN:

Retry Storm

A self-inflicted overload where many clients retry simultaneously after a failure and drive the system into the ground — effectively a DDoS against itself. A classic anti-pattern and common death spiral when AI agents loop requests without backoff on timeout or rate-limit, where each retry adds load instead of relieving it.
SEE ALSO:
DOMAIN:

Recall

How much of the relevant information a search system manages to find. High recall means the system misses fewer important hits.
SEE ALSO:
DOMAIN:

Feedback Loop

A process where user behavior or feedback improves the product over time. AI products often need strong feedback loops to become good in practice.
SEE ALSO:
DOMAIN:

Monitoring

Watching how the AI system behaves in production. This can cover errors, response quality, cost, latency, and user feedback.
SEE ALSO:
DOMAIN: