AI Glossary: 40 Terms Explained in Plain English
An AI glossary should do one job: let you look up a term and actually understand it, without needing to already know three other terms first. Most do not. Search for "what is a context window" and the top result defines it using "tokens" and "attention" as if those were common knowledge. This glossary is built the other way around, core concepts first, each one explained in isolation, so that by the time you reach the business and deployment terms near the end, every earlier term has already been defined in the plain language you actually think in.
Quick Reference: The 10 Terms You Will See Most Often
Before the full glossary, here are the ten terms that show up most often in everyday AI conversations and product documentation, defined in one line each.
| Term | One-Line Definition |
|---|---|
| Token | The actual chunk of text, often part of a word, that a model reads and predicts one at a time |
| Context window | The maximum amount of text a model can consider at once |
| Prompt | The instruction or question you give the model |
| Hallucination | A fluent, confident answer that is factually wrong |
| Inference | The model actually generating a response, as opposed to being trained |
| Fine-tuning | Additional training that shapes a model's behavior after its initial training |
| RAG | Retrieval-augmented generation: looking up documents before answering |
| Temperature | A setting controlling how predictable or varied a model's word choices are |
| Consensus Answer | A synthesised best answer combining and ranking multiple models' responses |
| Confidence score | A percentage reflecting how strongly multiple models agree on an answer |
Core Concepts
These ten terms form the foundation. Nearly every other term in this glossary builds on one of these.
How to Use This AI Glossary
Read the ten core concepts below first, since the model behavior, architecture, and business terms that follow all assume you already know these.
1. Token
The actual unit of text a language model processes. Often a whole common word, but just as often a fragment of a longer or rarer word, plus punctuation. Models read and generate text one token at a time, not one letter or one word at a time.
2. Context Window
The maximum amount of text, measured in tokens, a model can consider at once, including your prompt, uploaded files, and prior conversation. Content beyond that limit gets truncated or dropped.
3. Parameter
One of the billions of internal numeric values inside a model that were adjusted during training. Parameter count is often used as a rough, imperfect proxy for a model's overall capacity.
4. Inference
The act of a trained model actually generating a response to your prompt. Training happens once, before release; inference happens every single time you use the model.
5. Training
The process of repeatedly showing a model text, having it predict what comes next, and adjusting its parameters based on how close that prediction was, across a massive dataset.
6. Fine-tuning
A second, smaller round of training applied after the main training, used to shape a model's tone, behavior, or performance on a specific task without starting from scratch.
7. Prompt
The text you give a model, whether a question, an instruction, or both, that it uses as the starting point for generating a response.
8. Temperature
A setting that controls how predictable or varied a model's output is. Low temperature sticks to the most likely next token; high temperature allows more variety and creativity, at some cost to consistency.
9. Embedding
A numeric representation of text, an image, or other data that captures its meaning in a form a model can mathematically compare against other embeddings, which is how similarity search and retrieval work.
10. Latency
The time between sending a prompt and receiving a response. Lower latency means a faster-feeling tool, and it becomes especially noticeable when querying multiple models at once.
Model Behavior Terms
These ten terms describe how a model behaves once it is running, including why it sometimes gets things wrong.
11. Hallucination
A response that sounds fluent and confident but is factually incorrect, often an invented citation, statistic, or detail the model produced because it was statistically plausible, not because it was verified.
12. Grounding
Anchoring a model's response to a specific, verifiable source, such as a document or search result, instead of relying purely on what it memorized during training.
13. RAG (Retrieval-Augmented Generation)
A technique where a model looks up relevant documents or data before generating an answer, combining retrieval with generation to improve accuracy on specific, checkable facts.
14. Chain-of-Thought
A prompting or reasoning technique where a model works through intermediate steps explicitly before giving a final answer, which often improves accuracy on multi-step problems.
15. Alignment
The general effort to make a model's behavior match human intentions and values, covering everything from following instructions correctly to refusing harmful requests.
16. RLHF (Reinforcement Learning from Human Feedback)
A training technique where human reviewers rate different possible model responses, and the model is nudged toward the styles and behaviors that scored highly, shaping tone and helpfulness after initial training.
17. Bias
A systematic slant in a model's output, in framing, emphasis, or default perspective, that comes from its training data and human feedback tuning rather than a deliberate, disclosed choice.
18. Confidence Score
A percentage or rating reflecting how strongly a system believes an answer is correct, often based on agreement between multiple models or multiple attempts.
19. Zero-Shot and Few-Shot
Zero-shot means asking a model to perform a task with no examples given. Few-shot means giving it a small number of examples in the prompt to guide the expected format or style.
20. Guardrails
Rules, filters, or checks built around a model to prevent specific categories of unwanted output, such as harmful content or responses outside an intended scope.
Architecture and Technical Terms
These ten terms describe how models are built and deployed under the hood.
21. Transformer
The neural network architecture underlying almost every modern large language model, built around a mechanism called attention that lets the model weigh the relevance of different parts of the input to each other.
22. Attention
The mechanism inside a transformer that lets a model weigh how relevant each part of the input text is to every other part, which is central to how it captures context and meaning.
23. Neural Network
A computing structure loosely inspired by connections between neurons, made of layers of simple mathematical units that, adjusted together during training, can learn complex patterns from data.
24. Vector Database
A database designed to store and search embeddings efficiently, commonly used to power RAG systems that need to find the most relevant documents for a given query.
25. Multimodal
A model that can process and generate more than one type of content, such as text, images, and audio, rather than being limited to text alone.
26. Open-Weight Model
A model whose trained parameters are published so anyone can download, inspect, and run it independently of the original provider, unlike a closed model available only through a hosted API.
27. Quantization
A technique that reduces the precision of a model's parameters to make it smaller and faster to run, usually with a small tradeoff in output quality.
28. System Prompt
Instructions set behind the scenes, before your own prompt, that shape a model's overall behavior, tone, or constraints for the entire conversation.
29. API (Application Programming Interface)
A defined way for one piece of software to call another over the internet, used to send a prompt to a model and receive its response programmatically instead of through a chat interface.
30. Foundation Model
A large, general-purpose model trained on broad data, capable of being adapted through fine-tuning or prompting to many different downstream tasks rather than being built for just one.
Business and Deployment Terms
These final ten terms show up specifically when evaluating or deploying AI tools for a team, including several terms specific to how multi-model platforms like Talkory work.
31. LLM Orchestration
The practice of coordinating requests across multiple language models, routing, comparing, and combining their outputs, instead of relying on a single model for every task.
32. Model Router
A system that decides which model, or which set of models, should handle a given request, based on factors like task type, cost, or required capability.
33. SSO (Single Sign-On)
An authentication setup that lets users log into a tool using their existing company identity provider, instead of creating and managing a separate password.
34. Data Residency
A guarantee about which physical region or jurisdiction a company's data is stored and processed in, often required for regulatory or compliance reasons.
35. Rate Limit
A cap on how many requests an account or API key can make within a given time window, used to manage load and prevent abuse.
36. SLA (Service Level Agreement)
A formal, contractual commitment to a specific level of uptime, response time, or support, typically offered on enterprise plans.
37. Consensus Answer
A synthesised response built by combining and ranking multiple models' answers to the same question into one best final answer, rather than presenting a single model's output alone.
38. Common Answer
The specific points and ideas that appeared consistently across every model's response to the same question, surfaced separately from the synthesised Consensus Answer.
39. Recursive Correction
A process where a model reviews its own initial response, is asked to identify errors or gaps, and produces an improved second version, which is then combined with other models' corrected answers.
40. Wallet-Based Pricing
A pay-as-you-go billing model where a prepaid balance is deducted per query based on actual usage, instead of a fixed subscription regardless of how much you use the service.
See These Terms in Action
Run a query on Talkory and watch tokens, confidence scores, and Consensus Answers work in real time.
Try Talkory FreeTerms People Commonly Confuse
- Fine-tuning vs. prompting. Fine-tuning retrains the model itself; prompting only shapes a single conversation without changing the underlying model.
- Hallucination vs. bias. A hallucination is an invented, false detail; bias is a systematic slant in framing that can exist even when every stated fact is accurate.
- Token vs. word. A word is a unit of language; a token is the actual chunk of text the model processes, which does not always map one-to-one with words.
- Model vs. platform. A model, such as GPT or Claude, generates responses; a platform, such as Talkory, is the product that queries multiple models and compares their output.
- Confidence score vs. accuracy. A confidence score reflects how strongly models agree with each other, which correlates with but is not a guarantee of factual correctness.
“After testing multiple AI models on coding, research, and business prompts, combined outputs produced more reliable results than any single model.” Internal multi-model evaluation, Talkory research team.
Why Talkory Wins on Making These Terms Useful
Understanding this vocabulary matters most when you can actually see it working. Talkory queries GPT, Claude, Gemini, Grok, Perplexity Sonar, and Kimi K3 in parallel, so the terms in this glossary, token, context window, confidence score, Consensus Answer, Common Answer, Recursive Correction, are not abstract definitions but labels for things you watch happen on screen every time you run a query.
That turns a glossary from something you read once into a reference you can check against a real interface the next time a term shows up in a product update, an API response, or a conversation with a colleague.
Final Verdict: Vocabulary Is the Fastest Way to Stop Feeling Behind on AI
This AI glossary will not make you an engineer, and it does not need to. Most confusion about AI is not conceptual, it is vocabulary. Once you know what a token, a context window, and a hallucination actually are, headlines and product documentation stop sounding like a different language and start sounding like a system you can reason about.
The direct recommendation: bookmark this page rather than trying to memorize all 40 terms at once. Come back when a new term shows up in a tool you are using, and build the vocabulary the way it actually sticks, one real, specific use at a time.
Frequently Asked Questions
What is the difference between a token and a word in AI?
A word is a unit of human language. A token is the actual unit an AI model processes, which is often a whole common word but just as often a fragment of a longer or rarer word, plus punctuation and spacing. A single word can be one token or split across several, depending on how common it was in the model's training data.
What does context window mean in simple terms?
A context window is the maximum amount of text, measured in tokens, that a model can consider at once when generating a response, including your prompt, any uploaded documents, and the conversation history. Anything beyond that limit gets truncated or dropped, which is why very long conversations or documents can cause a model to lose track of earlier details.
What is RAG in AI and why does it matter?
RAG, retrieval-augmented generation, is a technique where a model looks up relevant documents or data before answering, instead of relying only on what it memorized during training. It matters because it lets a model answer using current or private information it was never trained on, and it usually improves factual accuracy on specific, checkable claims.
What does temperature mean when talking about AI models?
Temperature is a setting that controls how predictable or varied a model's word choices are. A low temperature makes the model stick closely to the single most likely next token, producing more consistent, conservative output. A high temperature allows less likely tokens to be chosen more often, producing more varied, creative, and occasionally less coherent output.
Why does this glossary matter for using AI tools like Talkory?
Knowing these terms helps you read AI-generated results, error messages, and API documentation accurately, and helps you understand what a confidence score, a Consensus Answer, or a hallucination actually means when a multi-model platform like Talkory shows it to you, rather than treating those outputs as an unexplained black box.
Get 5 AI perspectives on this topic
Talkory runs your question through GPT, Claude, Gemini, Grok, Sonar & Kimi K3 simultaneously, then cross-checks the answers.