AI & ML Tech Glossary

Agent evaluation measures how well an AI agent completes real tasks over multiple steps, including planning, tool use, reliability, safety, latency, and cost, not just...

Agent memory is the short- and long-term storage an AI agent uses to retain user preferences, past actions, and task context across steps or sessions,...

An agent runtime is the execution layer that runs an AI agent’s loop, manages state and memory, integrates tools, and enforces policies, logging, and guardrails...

An agent sandbox is a controlled, isolated environment that limits an AI agent’s permissions and tool access (files, network, APIs) to reduce risk, enable safe...

Chain-of-thought prompting is a technique that encourages an LLM to write intermediate reasoning steps before the final answer, which can improve multi-step problem solving and...

Constitutional AI is an alignment method that uses a written set of principles to guide an LLM’s self-critique and revision during training, helping it follow...

A context window is the token limit a language model can attend to in a single request. It determines how much conversation and document text...

Contextual compression shortens retrieved documents for RAG by extracting or summarizing only query-relevant information, reducing tokens and noise while preserving evidence for grounded LLM answers.

Direct Preference Optimization (DPO) aligns LLMs using preference pairs (chosen vs. rejected responses) with a direct training objective. It avoids training a separate reward model...

An embedding model converts text (or other data) into vectors where semantic similarity maps to vector proximity. These embeddings power semantic search and RAG by...

Evals are systematic tests for LLMs and LLM apps that score quality, safety, and reliability using datasets, rubrics, automated checks, and sometimes human review. They...