Learning LLMs & AI Agents

An AI system that uses an LLM as its reasoning engine to determine which actions to take and in what order, often interacting with external tools.

A deep learning model trained on vast text to predict the next token, enabling it to generate human-like text and perform reasoning tasks.

Retrieval-Augmented Generation — grounds an LLM's responses in external knowledge retrieved at runtime.

Model Context Protocol — an open standard enabling secure two-way connections between data sources and AI models.

Numerical representations of text in a high-dimensional vector space, capturing semantic meaning.

An AI system that uses an LLM to dynamically determine a sequence of actions, often utilizing external tools to achieve a goal.

A prompting technique that instructs the model to generate intermediate reasoning steps before arriving at a final answer, significantly improving logic performance.

The maximum number of tokens an LLM can process in a single prompt-response interaction.

High-dimensional vector representations of text. Text with similar semantic meaning will have vectors located close together in space.

Taking a pre-trained base model and training it further on a smaller, specific dataset to specialize its behavior or improve instruction adherence.

When an LLM generates a response that sounds plausible but is factually incorrect or unsupported by its training data or context.

A massive neural network trained on vast text corpora to predict next-token probabilities.

The practices and tools used to deploy, manage, evaluate, and scale LLM applications in production reliably.

An open standard protocol designed to securely connect AI models with external data sources and tools.

The iterative process of structuring text input to effectively communicate with and guide the outputs of generative models.

A security vulnerability where malicious input overrides system prompt instructions, causing the model to execute unintended behaviors.

Retrieval-Augmented Generation — grounding an LLM on external knowledge retrieved dynamically from a database to prevent hallucinations and access real-time data.

A foundational agent architecture combining "Reasoning" (thinking about what to do) and "Acting" (using a tool), iteratively.

A framework where an agent evaluates its own past actions and outcomes, generating verbal reinforcement to improve future attempts.

The ability of an LLM to output structured data specifying a function name and arguments, allowing the system to trigger external software.

An advanced reasoning technique extending CoT by allowing the model to explore multiple reasoning paths concurrently, evaluating and pruning them.

A specialized database optimized for storing, managing, and performing similarity searches on embedding vectors.