Prompt Engineering
The art and technique of crafting inputs to AI models in a way that produces the best possible responses. This includes techniques like few-shot learning, chain-of-thought, or setting system instructions. Prompt engineering is a core skill for working with modern language models.
RAG
Retrieval-Augmented Generation combines a language model with an external knowledge base. Before the model generates an answer, relevant information is retrieved from the database and provided as context. RAG reduces hallucinations and makes it possible to incorporate current or company-specific data.
Reinforcement Learning
A learning method where an AI agent learns to make optimal decisions through trial and reward. The agent receives positive or negative feedback for its actions and adjusts its behavior accordingly. Reinforcement learning is used for game strategies and fine-tuning language models, among other things.
Temperature
A parameter in text generation that controls how creative or random a language model's output is. A low temperature produces predictable, conservative answers, while a high temperature yields more creative but less reliable results. Temperature is typically set between 0 and 2.
Token
The smallest unit into which text is broken before a language model processes it. A token can be a whole word, a word fragment, or a single character. The token count determines how much text fits into a model's context window and influences usage costs.
Transformer
A neural network architecture introduced in 2017 that is based on the attention mechanism and enables parallel data processing. Transformers are the foundation of nearly all modern language models and have largely replaced older architectures like RNNs. The architecture has also proven successful in image and audio models.
Transfer Learning
An approach where a model pre-trained on a large dataset is reused for a new, related task. Instead of starting from scratch, the model leverages its already-learned knowledge as a starting point. Transfer learning saves significant time, computing power, and training data.
Vector Database
A specialized database that stores data points as high-dimensional vectors and can efficiently search for similar vectors. It is commonly used with embeddings to enable semantic search, recommendation systems, and RAG applications. Well-known examples include Pinecone, Weaviate, and Chroma.
Zero-Shot Learning
The ability of an AI model to solve a task without having been specifically trained on it or given examples. The model transfers its general knowledge to the new task. Zero-shot learning is especially impressive with large language models, which can solve many tasks from a prompt instruction alone.