AI & data glossary
A plain-language reference to the terms that come up when you build with AI and data - written for the business reader deciding what to build, not the researcher writing the paper. If a word in one of our articles is unfamiliar, it is probably defined here.
AI agent
A software system that uses an AI model to take actions toward a goal - reading data, calling tools, and making decisions - rather than just answering a single question. The useful ones operate inside guardrails and hand off to a human at the steps that carry risk.
Context window
The amount of text a language model can consider at once, measured in tokens. Everything the model "sees" for a request - your instructions, the retrieved data, the conversation so far - has to fit inside it, which is why large documents are usually chunked and retrieved selectively.
Embedding
A list of numbers that captures the meaning of a piece of text so a computer can compare it to other text by similarity rather than exact words. Embeddings are what let a system find the passage that answers a question even when it uses none of the same words.
ETL / ELT
Two patterns for moving data: Extract, Transform, Load versus Extract, Load, Transform. The difference is whether you clean and reshape the data before or after it lands in the destination. Both are about turning scattered source data into something dependable to query.
Evaluation (evals)
The practice of measuring an AI system against a fixed set of test cases so quality is a number you can track instead of a feeling. Good evals catch regressions before users do and are the difference between "it seemed fine" and "it passed the suite".
Fine-tuning
Further training a pre-trained model on your own examples so it adapts to a specific tone, format, or task. It is one of two main ways to specialize a model - the other is retrieval - and the right choice depends on whether you need new behavior or new knowledge.
Governance
The controls that make an AI system accountable: who is allowed to do what, which actions require approval, and how every decision is recorded. Governance is what lets an automation you trust also be one you can defend when someone asks what happened.
Guardrails
Constraints placed between a model and the systems it can affect - input validation, allowed-action lists, output checks - so a wrong or manipulated model output cannot do real damage. Guardrails belong at the action boundary, not only in the user interface.
Hallucination
When a language model produces confident, fluent text that is factually wrong or invented. It happens because the model predicts plausible language, not verified truth, which is why grounding answers in real data and citing sources matters.
Human-in-the-loop
A design where a person reviews and approves an AI system's proposed action before it takes effect. It is the standard pattern for any workflow where a mistake is costly - the model recommends, the human decides, and the decision is logged.
Inference
The act of running a trained model to get an output - the step that happens every time a user sends a request. Inference is where ongoing cost and latency live, so controlling it (model choice, caching, batching) is central to running AI affordably.
Large language model (LLM)
A model trained on vast amounts of text to predict and generate language. LLMs power chatbots, copilots, and agents, and can summarize, translate, classify, and write - but they reason over patterns in text, not a database of facts, which shapes how you must use them.
MLOps
The discipline of running machine-learning systems in production reliably: deployment, monitoring, versioning, and automated retraining. It is to ML what DevOps is to software - the difference between a model that works once and one that keeps working.
Prompt injection
An attack where untrusted data fed to a model contains hidden instructions that hijack its behavior - for example, text in a document that tells the model to ignore its rules. It is the AI equivalent of an injection vulnerability and is defended by treating external data as untrusted, never as instructions.
RAG (retrieval-augmented generation)
A technique that retrieves relevant passages from your own data and gives them to the model as context before it answers. RAG is how you get an LLM to answer from your documents accurately, with citations, without retraining it.
Token
The unit a language model reads and writes - roughly a word fragment. Models have token limits and are usually billed per token, so token count is both a technical constraint and the main lever on cost.
Vector database
A database built to store embeddings and find the most similar ones quickly. It is the retrieval engine behind RAG and semantic search - the component that, given a question, surfaces the handful of passages most likely to contain the answer.