Glossary of Prompt Engineering Terms
A comprehensive glossary of 100+ terms you'll encounter in prompt engineering.
Aβ
Agent - An AI system that can take actions autonomously, often using tools and making decisions.
Alignment - Ensuring AI behavior matches human intentions and values.
Anthropic - AI safety company, creator of Claude models.
API (Application Programming Interface) - Interface for programmatic access to AI models.
Attention Mechanism - Neural network component that weighs importance of different input parts.
Bβ
Batch Processing - Processing multiple prompts simultaneously.
Bias - Systematic errors in AI outputs, often reflecting training data prejudices.
BPE (Byte Pair Encoding) - Tokenization method used by many LLMs.
Cβ
Chain-of-Thought (CoT) - Prompting technique encouraging step-by-step reasoning.
ChatGPT - OpenAI's conversational AI product based on GPT models.
Claude - Anthropic's family of AI assistants.
Completion - The AI-generated text following a prompt.
Constitutional AI - Anthropic's approach to AI safety through principles.
Context - Background information provided to help AI understand the task.
Context Window - Maximum tokens an AI can process in one interaction.
Dβ
Data Extraction - Using AI to pull structured information from unstructured text.
Delimiter - Characters used to separate sections in prompts (e.g., """, ```).
Downstream Task - Specific application of a pre-trained model.
Eβ
Embedding - Numerical representation of text in vector space.
Emergent Abilities - Capabilities that appear unexpectedly in larger models.
Evaluation - Measuring AI output quality and effectiveness.
Fβ
Few-Shot Learning - Providing examples in prompts to guide AI behavior.
Fine-Tuning - Additional training of a model on specific data.
Foundation Model - Large model trained on broad data, adapted for specific uses.
Gβ
Generative AI - AI that creates new content (text, images, etc.).
GPT (Generative Pre-trained Transformer) - OpenAI's model architecture family.
Grounding - Connecting AI outputs to factual, verifiable information.
Hβ
Hallucination - When AI generates plausible but false information.
Hyperparameters - Settings that control model behavior (temperature, top_p, etc.).
Iβ
Inference - Using a trained model to generate outputs.
Instruction Tuning - Training models to follow human instructions.
In-Context Learning - AI learning from examples within the prompt itself.
Jβ
JSON Mode - Output format specification for structured data.
Jailbreaking - Attempting to bypass AI safety measures (unethical).
Kβ
Knowledge Cutoff - Date when model's training data ends.
Lβ
Llama - Meta's open-source LLM family.
LLM (Large Language Model) - AI models trained on vast text data.
LoRA (Low-Rank Adaptation) - Efficient fine-tuning method.
Mβ
Max Tokens - Setting limiting response length.
Mistral - French AI company and their open-source models.
Multimodal - AI handling multiple input types (text, images, audio).
Nβ
Natural Language Processing (NLP) - Field of AI dealing with human language.
Neural Network - Computing system inspired by biological neurons.
Oβ
One-Shot Learning - Providing a single example in a prompt.
OpenAI - AI company, creator of GPT models and ChatGPT.
Output Format - Specified structure for AI responses.
Pβ
Parameter - A value learned during model training (models have billions).
Persona - Role or character assigned to AI in prompts.
Prompt - Input given to an AI to generate a response.
Prompt Engineering - Art and science of designing effective AI inputs.
Prompt Injection - Attack attempting to manipulate AI behavior.
Qβ
Query - A request or question sent to an AI system.
Rβ
RAG (Retrieval-Augmented Generation) - Combining AI with information retrieval.
Reasoning - AI's ability to process logic and draw conclusions.
Retrieval - Fetching relevant information to augment AI responses.
Role Prompt - Assigning a specific persona or expertise to AI.
Sβ
Sampling - Process of selecting tokens during generation.
Self-Attention - Mechanism allowing models to weigh input relationships.
Semantic Search - Search based on meaning rather than keywords.
Stop Sequence - Characters that signal the model to stop generating.
System Prompt - Instructions given to the AI before user interaction.
Tβ
Temperature - Parameter controlling randomness in outputs (0-2, typically).
Token - Basic unit of text processed by AI (roughly 4 characters).
Tokenization - Breaking text into tokens for processing.
Top-p (Nucleus Sampling) - Parameter controlling diversity of outputs.
Transformer - Neural network architecture underlying modern LLMs.
Uβ
User Prompt - The actual request or question from the user.
Vβ
Validation - Checking AI outputs for accuracy and appropriateness.
Vector Database - Storage system for embeddings, used in RAG.
Verbosity - Amount of detail in AI responses.
Wβ
Workflow - Sequence of AI operations for complex tasks.
Xβ
XML Tags - Markup format especially effective with Claude models.
Zβ
Zero-Shot Learning - AI performing tasks without specific examples.
Quick Reference Tableβ
| Term | Brief Definition |
|---|---|
| CoT | Step-by-step reasoning prompting |
| Context Window | Max text AI can process |
| Few-Shot | Prompting with examples |
| Hallucination | AI generating false info |
| LLM | Large Language Model |
| RAG | AI + information retrieval |
| Temperature | Controls output randomness |
| Token | Basic text unit (~4 chars) |
| Zero-Shot | Prompting without examples |