AI glossary

A

Agent (AI agent)

A system that is given a goal, plans steps, uses tools such as search, files or email, observes the results and keeps going until it finishes or is stopped. It differs from a chat assistant, which returns text and leaves the next step to you.

In Keplar: Agents are available on Plus and above, with spend caps, approval for outbound actions and an audit log.

See: Agents overview, What is an AI agent?

AGI (artificial general intelligence)

A hypothetical AI that matches human ability across a broad range of intellectual tasks. It does not exist today, and experts disagree about how to define it or whether and when it might.

In Keplar: Keplar is not AGI and does not claim to be.

See: AI vs AGI vs superintelligence

AI (artificial intelligence)

A broad label for software that performs tasks normally associated with human intelligence, such as recognizing images, translating, predicting or generating text. Most modern systems learn patterns from data instead of following only hand-written rules.

See: What is AI?

Alignment

The work of making an AI system's behavior match the intentions and values of the people who build and use it. It covers following instructions, avoiding harmful output and not pursuing unintended goals.

See: AI vs AGI vs superintelligence

API (application programming interface)

A defined way for one program to request services from another. AI providers expose their models through APIs, so products can send a prompt and receive a response without running the model themselves.

In Keplar: Keplar calls providers' model APIs. It does not currently offer a public, documented API with keys.

See: Developers and API status

Approval gate

A step where a person must confirm an action before it happens. Typical for sending a message, publishing or paying.

In Keplar: Outbound agent actions and writes through connected apps wait for approval by default; scheduled runs always do.

See: Tool permissions and approvals

Attention

The mechanism inside a transformer that lets each part of the input weigh its relationship to every other part. It is a core reason modern language models handle context well.

See: Transformer

Audit log

A record of who or what did what, and when. For automated systems it lets you review actions after the fact.

In Keplar: Every owner decision, agent run, connector change, order and form submission is logged.

See: Agent guardrails

B

Benchmark

A standardized test used to compare models, such as a set of questions with known answers. Results depend on the exact test, prompt and date, and models can be tuned to do well on public benchmarks, so a score is a clue, not a verdict.

See: Best AI for coding

Bias

Systematic skew in a model's output caused by its training data, tuning or design. It can affect who or what is represented and how topics are framed.

C

Chain of thought

A way of getting a model to work through intermediate steps before answering. Written reasoning can improve results on multi-step problems, but it is not a guaranteed window into how the model actually arrived at its answer.

Chatbot

A program you converse with. Modern chatbots are usually built on large language models, though simple ones follow scripts.

Citation

A pointer to a source supporting a statement. AI-generated citations can be wrong or invented, so a citation is something to open, not proof.

In Keplar: Sources in a Keplar answer are labeled "Cited by <model>; Keplar didn't open or check".

See: Sources and citations

Classifier

A program that assigns an input to a category. In routing, a classifier decides what kind of question it is and how demanding.

In Keplar: A rule-based classifier sizes up each question by wording, length, attachments and topic.

See: Question understanding

Consensus

Agreement among several independent sources. Among AI models, agreement is useful evidence but not proof, because models can share training data and mistakes.

In Keplar: Shown as a level (consensus, partial, none, single, unscored) and as counts such as "4 of 5 models", never as a percentage of correctness.

See: Agreement and consensus, Can you trust AI consensus?

Context window

The amount of text a model can consider at one time, measured in tokens. It includes your prompt, the earlier conversation, any documents and the reply being written.

See: AI memory and context windows

Credit

A unit some products use to meter how much AI work you consume, so heavier tasks cost more than lighter ones.

In Keplar: Keplar Usage is measured in credits derived from real provider charges, shown to you as a meter and a level.

See: Credits explained

D

Deep research

A mode in which an AI system plans sub-questions, searches or reads many sources over minutes and writes a longer report, usually with links. The report still needs checking.

In Keplar: Available on Plus and above; a typical run is estimated at about 1,200 credits.

See: Deep Research, What is deep research?

Diffusion model

A kind of generative model, widely used for images, that learns to create a picture by gradually removing noise from a random starting point, guided by a text description.

Dissent

A minority position that differs from the majority. Keeping dissent visible helps, because the minority is sometimes the one that is right.

In Keplar: Kept as a labeled dissent in the answer when a reviewer judges it well supported.

See: Disagreement detection

E

Embedding

A list of numbers that represents the meaning of a piece of text, image or other data, so similar items sit close together. Used for search and retrieval.

Ensemble

A method that combines the outputs of several models to get a better or more reliable result than any one alone.

See: What is multi-model AI?

Evaluation (eval)

A repeatable test of how well a model or system performs on tasks you care about. A good personal eval uses your own real tasks.

See: Best AI for coding

Exact check

Computing an answer with code or a deterministic method instead of letting a model guess, used for arithmetic, counting and similar precise questions.

In Keplar: Arithmetic, counting and exact-fact questions are checked by code, and the answer is built around the result.

See: Exact checks

F

Fine-tuning

Continuing the training of an existing model on additional examples to specialize its behavior or style.

Foundation model

A large model trained on broad data that can be adapted to many tasks, such as a large language model or a large image model.

G

Generative AI

AI that produces new content, such as text, images, audio, video or code, instead of only classifying or predicting.

See: What is AI?

GEO (generative engine optimization)

Efforts to make content more likely to be found, used and cited by AI assistants and AI search features. Overlaps heavily with classic SEO; guarantees do not exist.

See: How to show up in AI answers

Grounding

Anchoring a model's answer in supplied or retrieved material, such as a document or search results, to reduce unsupported claims. Grounded answers can still misread their material.

Guardrails

Limits and checks around an AI system, such as spending caps, permission rules, content filters and approval steps.

In Keplar: Agents have caps, approval gates, a kill switch and an audit log.

See: Agent guardrails

H

Hallucination

Output that is fluent and plausible but false, such as an invented citation or fact. It happens because language models generate likely text rather than looking facts up.

See: What are AI hallucinations?

I

Inference

Running a trained model to produce an output for a new input. Inference is what happens each time you ask a question, and it is where most ongoing cost lies.

J

Judge model

A model asked to evaluate or choose between other outputs, or to write a final synthesis from them.

In Keplar: A judge-tier model breaks ties and writes the synthesis; for the Free plan no judge tier is used.

See: Synthesis

K

Knowledge cutoff

The date after which a model's training data contains little or nothing. Without a live lookup tool, a model may not know about later events.

In Keplar: Normal Keplar questions do not use live web lookups.

See: What Keplar cannot do

L

Large language model (LLM)

A neural network trained on very large amounts of text to predict the next piece of text. After tuning, it can answer questions, write, summarize and generate code.

See: What is AI?

Latency

How long a response takes to start or finish. Panels of models are limited by their slower members unless stragglers are cut off.

In Keplar: A whole answer has a time budget of about 60 seconds, and slow models are not waited for beyond a short grace.

See: Missing models and timeouts

M

MCP (Model Context Protocol)

An open protocol for connecting AI applications to external tools and data through servers that expose tools, resources and prompts in a standard way.

In Keplar: Keplar connects outbound to remote MCP servers on Plus and above, with per-tool permissions.

See: What is MCP?, Connected apps overview

Memory (AI memory)

A feature that saves facts about you between conversations and supplies relevant ones later. Distinct from the context window, which is only the current conversation.

In Keplar: On by default for signed-in accounts and fully controllable: view, add, delete or turn off.

See: Memory

Mixture of agents

An approach in which several models draft answers and another model reads those drafts to produce a refined one.

See: What is multi-model AI?

Mixture of experts (MoE)

A model architecture that contains many specialized sub-networks and activates only some for each token. It is a design inside one model, not several separate models giving opinions.

Model router

Software that decides which model should handle a request, usually to balance quality, speed and cost.

In Keplar: Keplar scores models by quality, reliability and latency divided by cost, then seats a panel.

See: Routing and panels

Multi-model AI

Any system that uses more than one AI model for a task, whether to route, fall back, vote, compare or synthesize.

See: Multi-model AI

Multimodal

Able to work with more than one kind of input or output, such as text, images, audio and video.

In Keplar: Questions with images or video go to vision-capable seats.

See: Attachments, images and video

N

Neural network

A computing system loosely inspired by the brain, made of layers of simple units whose connection strengths are adjusted during training.

O

Open-weight model

A model whose trained parameters are published so others can run or adapt it. Not always the same as open source, because licenses and training data may be restricted.

In Keplar: Panels include an open-weight seat where possible, for diversity and cost.

See: Model diversity

P

Panel

The set of models consulted for a single question. Each member answers independently, and the choice of members matters as much as the number.

In Keplar: Up to six models, varying by the question and your plan; Free is capped at three.

See: Routing and panels

Parameter

A learned number inside a model. Model sizes are often described by parameter count, though more parameters do not always mean better results.

Prompt

The input you give a model: a question, instruction, examples or documents.

See: How to write better AI prompts, Prompt library

Prompt injection

An attack in which text in content the model reads, such as a web page or email, contains instructions that try to override what the user or developer intended.

In Keplar: Tool output is treated as data: stripped of instruction-like lines, fenced and capped, and unable to change permissions.

See: Connected app security

R

RAG (retrieval-augmented generation)

A pattern in which a system first retrieves relevant documents, then gives them to a model to write an answer, so the answer can draw on material outside the model's training.

Reasoning model

A model trained or configured to spend extra computation working through a problem step by step before answering. Often better at multi-step tasks, slower and costlier.

RLHF (reinforcement learning from human feedback)

A tuning method in which people rate model outputs and the model is adjusted toward higher-rated behavior. It can improve helpfulness and also encourage agreeable answers.

See: Why does AI agree with you?

S

Sampling and temperature

Models pick each next token from a probability distribution. Temperature is a setting that controls how adventurous that choice is: low is more predictable, high is more varied.

See: Why AI models disagree

Superintelligence

A hypothetical AI that greatly exceeds the best human minds in virtually every field. It is a topic of research and debate, not an existing product.

In Keplar: Keplar is not superintelligence and does not claim to be.

See: Superintelligence

Sycophancy

A model's tendency to match what the user appears to believe or want instead of what is most accurate.

See: Why does AI agree with everything you say?

Synthesis

Combining several responses into one answer that keeps the points they share and explains where they differ.

In Keplar: A judge-tier model writes it, weighing evidence rather than headcount.

See: Synthesis

System prompt

Hidden instructions a product gives a model before your message, setting its role, style and limits. It differs between products and is one reason the same model can behave differently in different apps.

T

Token

A piece of text, often part of a word, that models read and write. Limits and prices are commonly measured in tokens.

Tool use (function calling)

Letting a model request that a program run an action, such as a search, a calculation or an API call, and then use the result.

See: Use connected apps in chat

Training data

The text, images, code and other material a model learns from. Its coverage, quality and date range shape what the model knows and gets wrong.

Transformer

The neural network architecture, introduced in 2017, behind most current language models. It uses attention to relate every part of the input to every other part.

V

Verification

Checking a draft answer against evidence. The strength of the check depends on what is examined.

In Keplar: A reviewer reads the draft against the models' responses; Keplar does not open or check sources.

See: Verification and review

Vision model

A model that can interpret images or video frames, not only text.

W

Web grounding (live search)

Letting a model run web searches and use the results when answering, which helps with recent information and also introduces the quality problems of web pages.

In Keplar: Normal Keplar questions do not use live web lookups; Deep Research gathers sources.

See: What Keplar cannot do

Z

Zero-shot and few-shot

Zero-shot means asking a model to do a task with no examples; few-shot means including a few examples in the prompt to show what you want.