Clear definitions for the AI terms buyers, builders, and operators run into every week. Use this page to get aligned on vocabulary before choosing tools, models, or workflows.
An AI agent is a system that can decide what step to take next, use tools, and keep moving toward a goal instead of answering only one prompt at a time.
Teams evaluating agent products need to separate simple chat wrappers from systems that can actually plan, retrieve data, and execute tasks.
A support agent might look up order data, draft a response, and escalate edge cases without a human manually switching tools.
An agentic workflow is a multi-step process where an AI system plans, calls tools, checks results, and iterates toward an outcome.
This is the practical difference between one-off content generation and automation that can handle more complex business processes.
A coding workflow can inspect a repo, edit files, run checks, and revise based on failures.
A context window is the amount of text, code, or multimodal input a model can consider in one request.
Context limits determine whether a model can handle large contracts, research packets, or codebases without splitting the work up.
If your prompt and source material exceed the model context, quality drops or the request fails outright.
A token is a small unit of text that models use for reading and generating content. Pricing and context limits are usually measured in tokens.
Token counts directly affect API cost, latency, and whether a workload fits into a model at all.
A long prompt with a large knowledge base chunk can become expensive even before the model writes any output.
LLM stands for large language model, a system trained on massive text or code corpora to predict and generate language.
Most AI buying decisions start with model choice because model quality, cost, speed, and tool access shape the rest of the stack.
Teams often compare frontier models for reasoning and cheaper models for background automation.
A multimodal model can work across more than one input or output type, such as text, images, audio, or video.
Multimodal capability matters when you need one system to interpret screenshots, documents, voice, or visual creative assets.
A multimodal workflow might read a PDF, inspect charts, and answer questions in plain language.
RAG stands for retrieval-augmented generation. It means fetching relevant external information before the model writes an answer.
RAG is one of the most common ways to make AI systems more useful on company-specific knowledge without retraining a model.
A team can connect docs, help center articles, or internal policies so the model answers from fresh source material.
A vector database stores embeddings so a system can retrieve semantically similar content rather than matching only exact keywords.
Vector retrieval is foundational for many RAG systems, especially when knowledge is spread across large document collections.
Instead of searching for exact phrasing, the system can find passages that are conceptually similar to the user question.
Fine-tuning means training a base model further on a specialized dataset so it behaves better for a narrow job or domain.
It can improve consistency for repeated tasks, but it is usually more expensive and slower to maintain than a prompt-plus-RAG approach.
A company might fine-tune a model for niche classifications after prompt engineering stops improving accuracy.
Prompt engineering is the practice of structuring instructions, examples, context, and constraints so a model produces more reliable output.
For most teams, better prompts are the fastest and cheapest performance improvement available.
Adding role, task, format, and examples often lifts quality before you need model changes.
A hallucination is when a model states something false, unsupported, or invented as if it were correct.
Hallucination risk shapes where AI can be safely automated and where human review or grounded retrieval is still required.
Models may invent sources, quote nonexistent policies, or confidently misstate numbers when no trusted reference is supplied.
Grounding means tying a model response to verified sources, live system data, or retrieved documents instead of relying only on model memory.
Grounded outputs are usually easier to audit, safer for production, and more defensible in buyer-facing or regulated workflows.
A finance workflow can ground a response in policy docs and ERP data before drafting an answer.
A benchmark is a standardized test used to compare model performance on tasks such as reasoning, coding, math, or factual recall.
Benchmarks can help narrow a shortlist, but they rarely map perfectly to your real workflow or internal data.
A model that leads on public reasoning tests may still underperform on your support tickets or pricing extraction workflow.
Inference is the act of running a trained model to generate an output from a given input.
Inference cost and latency become central once you move from experimentation to production traffic.
A cheap model may still be the wrong choice if repeated inference calls and retries create hidden cost.
MCP stands for Model Context Protocol, a standard way for models or agents to connect to tools, data sources, and external systems.
Standardized tool connections reduce brittle one-off integrations and make agent systems easier to extend.
An MCP server can expose CRM data, ticket history, or internal docs to an AI assistant through a consistent interface.
Move from definitions to decisions with the live pricing table, model comparison tool, prompt library, and stack finder.
Start with LLMs, tokens, context windows, prompt engineering, RAG, hallucination, and AI agents. Those terms directly affect cost, quality, and implementation decisions.
RAG brings fresh external information into a response at runtime, while fine-tuning changes the model itself with additional training data. Most teams try RAG first because it is easier to update and cheaper to maintain.
They determine whether your workload fits into a model request and how much each request costs. Those two constraints shape everything from product UX to operating margin.