Glossary
Agentic AI
AI systems that can work towards a goal by planning and carrying out multiple steps, rather than only responding to a single prompt. Agentic AI may use tools, retrieve information, make intermediate decisions and adapt its actions based on what happens during a task.
Agents
AI systems or configurations designed to carry out tasks with a degree of autonomy. An agent can combine a language model with instructions, information and tools to perform several actions towards a particular goal. The amount of autonomy and the actions available to an agent can vary considerably.
API (Application Programming Interface)
Like a power strip for software - lets different programs communicate. Extensions use this.
Bias
When the AI (unintentionally) adopts biases from its training data, which can result in unfair or skewed answers.
Context engineering
The deliberate process of selecting, structuring and combining the information an AI model receives so that it can perform a task effectively. Context can include prompts, instructions, documents, previous conversations, data and other types of input. Good context engineering helps produce more relevant, consistent and grounded responses.
Context Window
The 'working memory' of the AI - the maximum amount of text (tokens) the model can process at one time.
Extension
An extra tool or plug-in that gives UvA AI Chat new features, often by linking with other systems via an API.
Extensions
Additional tools or connections that extend what an AI model can do beyond generating a response from the conversation alone. Extensions can, for example, allow AI to retrieve information from another source, interact with a service or perform a specialised task.
Hallucination
When the AI presents nonsense or wrong info that sounds convincing. Always double-check.
Knowledge
Information made available to an AI system as background or reference material for its responses. In UvA AI Chat, knowledge can include uploaded files, project information and stored knowledge cards. Providing relevant knowledge can help ground responses in specific sources, but the AI can still interpret this information incorrectly.
Legacy models
Older AI models that remain available for compatibility or specific use cases after newer models have been introduced. Legacy models may differ from current models in their capabilities, speed, efficiency or supported features.
LLM (Large Language Model)
A very large AI model trained on massive amounts of text, capable of understanding and generating human language. This is the engine behind the chat.
Memory
A feature that allows information from earlier interactions to remain available in later conversations. In UvA AI Chat projects, Memory can connect files, instructions, previous conversations and knowledge cards across multiple chats. Memory is different from a model being retrained: stored information is supplied as context when relevant.
Model selection
Choosing which available AI model to use for a particular task. Different models can vary in capabilities, speed, energy use and suitability for tasks such as writing, reasoning, coding or working with different forms of input. Selecting a model therefore involves matching the model to the needs of the task.
Multimodal
The ability of an AI system to work with more than one type of input or output, such as text, images, audio, documents or other data. A multimodal model can combine information from different formats within the same task, depending on the capabilities available in UvA AI Chat.
Persona
A set of instructions that gives the AI a specific role or style (e.g., expert, coach).
Prompt
Your instruction, question, or text for the AI. The foundation of everything.
Prompt Engineering
The art of writing effective prompts to get exactly the output you want.
Token
The smallest unit of text for the AI (a word, or part of a word). The "context window" is measured in tokens.