Reference
AI Glossary
Clear definitions for the vocabulary behind modern AI products and research.
A
- Agent
An AI system that plans and uses tools (browsers, code, APIs) to complete multi-step goals.
C
- Context Window
The maximum amount of tokens a model can consider in one request, including prompt and output.
E
- Embeddings
Numeric vector representations of text or media that capture semantic similarity for search and clustering.
- Evaluation (Evals)
Systematic measurement of model quality using benchmarks, human ratings, or automated graders.
F
- Few-Shot Learning
Providing a handful of examples in the prompt so the model can mimic a pattern without fine-tuning.
- Fine-Tuning
Further training a pretrained model on domain-specific data to specialize behavior or style.
G
- Guardrails
Policies, filters, and classifiers that constrain unsafe or off-brand model behavior.
H
- Hallucination
When a model produces fluent but incorrect or fabricated information presented as fact.
I
- Inference
Running a trained model to generate outputs; distinct from training or fine-tuning.
L
- Large Language Model (LLM)
A neural network trained on massive text corpora to predict and generate language, powering chatbots and copilots.
- Latency
Time delay between sending a request and receiving a model response—critical for UX.
- LoRA
Low-Rank Adaptation—an efficient fine-tuning method that updates small adapter weights instead of the full model.
M
- Multimodal Model
A model that accepts or generates multiple modalities such as text, images, audio, and video.
O
- Open-Weight Model
A model whose weights are publicly downloadable for local or self-hosted inference.
P
- Prompt Engineering
The practice of designing instructions, examples, and constraints to steer model outputs reliably.
R
- Retrieval-Augmented Generation (RAG)
Architecture that retrieves relevant documents and feeds them to an LLM to ground answers in private or fresh data.
S
- Streaming
Delivering model tokens to the client as they are generated instead of waiting for the full response.
- System Prompt
Hidden or privileged instructions that set role, safety rules, and default behavior for an assistant.
T
- Temperature
Sampling parameter controlling randomness; lower values are more deterministic, higher more creative.
- Tokens
Units of text (subwords/characters) that models process; pricing and context limits are measured in tokens.
- Tool Calling
Capability allowing models to invoke functions or APIs with structured arguments during a conversation.
- Transformer
Neural architecture using attention mechanisms that underpins most modern LLMs and multimodal models.
V
- Vector Database
A database optimized for storing embeddings and performing nearest-neighbor similarity search.
Z
- Zero-Shot
Asking a model to perform a task with instructions only—no examples in the prompt.