Glossary · 24 terms

GenAI

Prompts, RAG, agents and how models run.

Beginner · Students and graduates

Prompt: The input text a model responds to. Clear goals, context and examples in it usually matter more than clever wording.
Token: Models split text into tokens, often parts of words. Limits and pricing are counted in tokens, not characters.
LLM: A large language model: a neural network trained on text that generates language by predicting likely next tokens.
Model: The learned parameters and architecture that produce predictions. Different models trade off quality, speed and cost.
Chatbot: An app that holds a back-and-forth conversation, sending the chat history to a model on each turn.
Dataset: The data used to train or evaluate a model. Its quality and coverage shape what the model can and can’t do.
Bias: Systematic skew in outputs, often reflecting imbalances in training data. It needs testing and mitigation, not just good intentions.
Context: The context window holds the prompt, history and any added documents. Anything outside it, the model simply doesn’t know.

Embedding: A vector representation where similar meanings sit close together. It powers semantic search and clustering.
RAG: Retrieval-augmented generation: search your own data, put the best matches in the context, and have the model answer from them.
Agent: A system where the model decides which tool to use, sees the result and continues until the task is done.
Fine-tune: Adjusting a pre-trained model’s weights with task-specific data to change its style or behaviour.
Vector DB: Stores vectors and finds the nearest ones to a query quickly, usually with approximate nearest-neighbour indexes.
Grounding: Supplying facts in the context and asking for answers based on them, often with citations, to reduce made-up content.
Tool use: The model returns a structured request to call a tool you defined; your code runs it and sends back the result.
Chunking: Breaking text into sections sized for embedding and retrieval. Chunk size and overlap strongly affect RAG quality.

Attention: Self-attention lets each token look at every other token and decide how much each matters. It’s the core of transformers. Docs
MCP: The Model Context Protocol standardises how applications expose tools, resources and prompts to AI models. Docs
Evals: Sets of inputs with expected outcomes or graders, run on every change so you know whether quality went up or down.
LoRA: Low-Rank Adaptation freezes the base weights and learns small low-rank updates, making fine-tuning far cheaper. Docs
Quantize: Quantization reduces numeric precision, such as 16-bit to 4-bit, cutting memory and speeding up inference with some quality loss.
Reranker: After a fast retrieval step, a reranker scores each query and document pair more precisely and keeps the best few.
Guardrail: Validation, filters and policies applied before and after model calls, such as schema checks or content rules.
Inference: The prediction phase, as opposed to training. Latency, throughput and cost per token are inference concerns.