A

AI Agent

A system where a language model decides which actions to take, uses tools, observes results and repeats until a task is done.

B

BM25

A classic keyword-ranking algorithm used by search engines that scores documents by term frequency and rarity.

C

Chain of Thought

Prompting a model to reason through intermediate steps before answering, which improves accuracy on multi-step problems.

Chunking

Splitting documents into smaller passages before embedding them, so retrieval can find the specific part that answers a question.

Context Window

The maximum number of tokens a model can consider at once, including both the input prompt and the generated output.

Cosine Similarity

A measure of how similar two vectors are, based on the angle between them: 1 means same direction, 0 unrelated, −1 opposite.

E

Embedding

A list of numbers (a vector) that represents the meaning of a text, so similar meanings end up close together.

Evals (LLM Evaluation)

Tests for AI behavior: datasets of realistic inputs plus graders that score outputs, run on every change.

F

Few-Shot Prompting

Including a few input→output examples in a prompt so the model learns the desired format and behavior.

Fine-Tuning

Further training a pretrained model on your own examples to change its behavior, style or performance on a narrow task.

Function Calling (Tool Use)

A model capability where it requests that your code run a named function with JSON arguments, then uses the result.

G

Gradient Descent

The optimization method that trains neural networks by repeatedly nudging parameters opposite to the gradient of the loss.

Guardrails

Checks and limits around an AI system, such as input filters, output validation, permissions and approvals, that keep it safe and on-task.

H

Hallucination

When an AI model produces fluent, confident output that is false or not supported by its sources.

Hybrid Search

Combining keyword search (like BM25) with vector search to get both exact-term matching and meaning-based matching.

L

Large Language Model (LLM)

A neural network trained on huge amounts of text to predict the next token, which lets it write, summarize, answer questions and write code.

Latency (TTFT)

How long an AI request takes. Time to first token (TTFT) measures when output starts, and total time includes generating every token.

LLM-as-Judge

Using a language model to grade another model's outputs against a rubric.

Logits

The raw, unnormalized scores a model outputs for each possible next token, before softmax turns them into probabilities.

LoRA (Low-Rank Adaptation)

A parameter-efficient fine-tuning method that trains small low-rank adapter matrices instead of all of a model's weights.

M

MCP (Model Context Protocol)

An open standard, introduced by Anthropic in 2024, for connecting AI applications to tools and data through reusable servers.

Multi-Agent System

An AI system where several model instances with separate roles and contexts work together, such as a lead agent and subagents.

N

Neural Network

A model made of layers of simple units that compute weighted sums followed by nonlinear activations, trained with gradient descent.

O

Overfitting

When a model memorizes its training data and performs much worse on new data.

P

Pretraining

The first, largest training stage, where a model learns language and knowledge by predicting the next token on huge text datasets.

Prompt Caching

A provider feature that reuses the processing of a repeated prompt prefix, cutting the cost and latency of those tokens.

Prompt Engineering

Designing the instructions, examples and context given to a model so it produces the output you need reliably.

Prompt Injection

An attack where text the model reads, from a user or hidden in content, overrides its instructions.

R

RAG (Retrieval-Augmented Generation)

A technique that retrieves relevant documents at question time and gives them to the model, so it answers from your data with citations.

Reasoning Model

A language model trained to think through problems step by step before answering, trading speed and cost for accuracy.

Reranking

Re-scoring a shortlist of retrieved results with a more accurate model (often a cross-encoder) to put the best ones on top.

RLHF

Reinforcement learning from human feedback: training a model toward answers people prefer, using human rankings of outputs.

S

Self-Consistency

Sampling several reasoning paths for the same question and taking the most common answer to improve accuracy.

Softmax

A function that turns a list of scores (logits) into probabilities that are all positive and sum to 1.

Streaming

Sending a model's response in small pieces as it's generated, usually via Server-Sent Events, so users see text immediately.

Structured Output

Getting a model to return machine-readable data, usually JSON that matches a schema, instead of free text.

System Prompt

Instructions that apply to an entire conversation, like role, rules, tone and output format, sent separately from user messages.

T

Temperature

A sampling setting that controls randomness: low values make output focused and repeatable, high values make it more varied.

Token

The unit of text a language model reads and writes: a common word, part of a word, a punctuation mark or a space.

Tokenizer

The component that splits text into tokens and maps each one to an integer ID before the model processes it.

Top-p (Nucleus Sampling)

A sampling method that picks only from the smallest set of tokens whose probabilities add up to p.

Transformer

The neural network architecture behind modern LLMs, built from stacked attention and feed-forward layers.

V

Vector Database

A database designed to store embeddings and quickly find the vectors most similar to a query.