Kamus AI

Honesty about who did what in learning work, including the parts AI helped with.

Checkable conditions for calling a feature done.

Your own code that runs automatically at set points in an agent loop, for example before a tool is used.

An agent's repeating cycle: the model picks a step, the app runs a tool, the result goes back to the model.

A packaged procedure (instructions, examples, scripts) that an agent loads only when the task fits.

A model that can use tools and take several steps on its own to reach a goal.

A closed list of actions or commands that may run; anything not listed is refused.

Processing data until there is no reasonable way left to identify the person.

The door a service provides so other programs can call it under clear rules.

A secret that proves a program may use a service, usually tied to a bill.

A list of criteria and performance levels used to assess work consistently.

A lasting record of who did what, when, to which data.

A decision about a person taken by a system without meaningful human involvement.

A small piece of code that runs part of an app and checks the result itself, every time it runs.

Making repeated work run by itself without a person handling it each time.

A separate line of work in a repository: its changes do not touch the main version until merged.

The step that turns code into a ready-to-run app, catching many mistakes on the way.

A temporary copy of a result, kept so the next request is faster.

A conversation with AI through an assistant app: you write, the AI answers, and the conversation can go on.

A list of steps or conditions ticked off one by one before work counts as done.

Personal data about a child; the PDP Law classes it as specific personal data.

Reading a code change before it is merged, to catch mistakes, gaps and unproven claims.

An AI agent that reads, writes and runs code in a repository, beyond suggesting snippets.

One set of changes saved to the git history, with a message that explains why.

Summarising a long conversation to fit the context window, at the risk of losing important details.

A link that gives an AI app access to another service, such as email, a calendar or file storage.

A ready-to-run package holding an app and everything it needs, built once and run anywhere.

The background you give so the model knows the situation, goal and limits of a task.

Deciding what goes into a model's context window: instructions, files, examples and notes, in the right amount.

The maximum number of tokens a model can see in one go.

Sending personal data to a party or server outside Indonesia.

A customer's clear statement that they agree to their data being used for a particular purpose, such as receiving promotions.

How long data is kept, and when it is deleted or destroyed.

Where an app keeps its data in order, so it can be searched, changed and access-controlled.

A numbered script that changes a database schema step by step, run in the same order in every environment.

Finding the cause of an error or wrong behaviour, then proving the fix with a test.

Opening and closing markers in a prompt that separate material from instructions, such as DATA START and DATA END.

A library or package written by others that your project uses, with its version.

A deploy installs a new version on a server; a release is that version once people start using it.

The list of lines added, changed or removed between two versions of a file.

Data made up on purpose for trying or teaching, with no real people, figures or documents.

A list of numbers that stands for a text's meaning, so similar texts get nearby numbers.

An address on a server that an app calls for one kind of request.

A setting given to an app from outside the code, such as a database address or a feature flag.

A sign that a program failed to do something, usually with a message that points at the cause.

A measured way to judge whether AI output is good enough for its task.

One or more sample inputs and outputs placed in a prompt so the model copies the pattern.

An ordered record of every agent step: messages, tool calls, inputs, results, errors and cost.

Kinds of information not opened to requesters under Article 17 of the Public Information Disclosure Law.

The number a command reports when it ends: usually 0 means success, anything else means failure.

A setting that turns a feature on or off without releasing new code.

A routine for checking figures: recompute, cross-foot and compare with the source.

Further training of an existing model on specific examples so it behaves better for a given task.

A ready-made structure that sets how an app is organised, so you only fill in your part.

A limit placed around an AI or agent so its mistakes cannot cause damage.

Model output that sounds convincing but is false or has no basis.

Data and information about a person's health; under the PDP Law it is specific personal data.

A periodic signal a process sends to show it is still running.

The principle that important AI decisions must be checked and approved by a person.

A unique marker for one request, so a repeated request is not carried out or charged twice.

A write-up after something went wrong: what happened, its impact, its cause, and the guardrail added.

A program trained on a very large amount of text to predict how text continues.

The lawful reason that allows personal data to be processed under the PDP Law.

The principle of giving an agent or a person only the permissions their task needs, no more.

Records a program writes as it runs: what happened, when, and with what result.

A way of making programs learn patterns from example data, instead of writing every rule by hand.

An open protocol that standardises how AI apps connect to tools and data sources.

A document holding a patient's identity, examinations, treatment, procedures and other services given to them.

A feature that makes an AI app remember things about you beyond one conversation.

Combining the changes of one branch into another, usually into the main branch.

Recording every paid AI use, per user or per feature, so costs can be capped.

A fake stand-in for a service or data in a test, so the test is fast and leaves real systems alone.

A model's tendency to produce output that is unfair or skewed for certain groups of people.

The company that runs an AI model and sells access to it, through an app or an API.

The process of shaping a model from large amounts of data before it is used.

A model that can work with more than one kind of input, such as text, images and audio.

Data born from daily work: deliveries, trips, stock, schedules and invoices.

A figure far from the rest of the data's pattern, such as a delivery cost ten times the usual.

The maximum number of tokens a model may write in one answer.

The tool that installs, updates and pins your project's dependencies.

A small change that fixes a bug or closes a vulnerability in a version already running.

Indonesia's Law Number 27 of 2022 on Personal Data Protection.

Data about a person who is identified or can be identified.

A chain of automatic steps that runs the build, tests and deploy on every change.

Writing so readers can find, understand and use information in one reading.

Production is the version real people use; staging is a copy for trying things before they go there.

A written file where an agent programme records each step, decision and status.

A space in an AI app that keeps instructions, files and conversations for one piece of work together.

The instruction you write for an AI: what you want, for whom, from which material.

An attack in which text slips in orders to steer a model off course.

A collection of tested, ready-to-use prompts with placeholders you fill in.

A rough version of a tool or page, made quickly to test with users before serious building.

Replacing direct identifiers such as names with codes; the data can still be traced back to the person.

Information produced, stored or managed by a public body and open to requesters, except what is exempt.

Sending commits from your laptop to the shared repository, so other people and automated systems can see them.

A request to a database to read, add, change or delete data.

A way to give a model relevant document passages before it answers, so the answer rests on sources.

A model that writes hidden thinking steps before giving its final answer.

Matching two records of the same transactions, then explaining every difference.

A project folder whose change history is tracked by git, on your laptop or on a shared server.

Going back to the version that last worked after a new release goes wrong.

A record of tested operating steps, written so someone else can repeat them safely.

The environment that runs your code, with its version, such as the Node or Python version on a server.

A separate workspace where an agent may get things wrong without touching data or systems people use.

A database's structure: tables, columns, data types and the rules that bind them.

A set of libraries and tools from a provider for calling its service from your code.

A password, access token or API key: anything that grants access and must not be seen by others.

A flaw in software that someone else can use to do something they should not be able to do.

A computer that runs an app or service and answers requests from other computers.

A pointer to the document a claim comes from, so people can check it.

A short write-up of what is built, for whom, what is out of scope, and when it counts as done.

A table file of rows, columns and formulas, like a workbook in a spreadsheet app.

An agent started by another agent for one part of a task, with its own context and permissions.

Standing instructions that apply to a whole conversation, outside the user's messages.

A list of candidates not hired now but kept, with consent, for future openings.

A setting that controls how random the model's word choices are.

The window where you type commands to the computer; a CLI is a program run from there.

A fixed set of examples with expected answers, for testing a prompt or model again and again.

Writing the test before the code, then watching it fail for the right reason.

A piece of text a model reads and writes, usually part of a word.

A model's ability to ask an app to run a given function, such as a search or a calculation.

An automatic check that the data your code uses matches its declared types, before the program runs.

A system that records every change to a project's files, who made it and why, so changes can be compared and undone.

A process running behind the app to handle queued jobs, such as sending email or calling AI.

A fixed sequence of steps, some done by AI, in an order the AI does not change on its own.

An extra working copy of one git repository, with its own branch, in a separate folder.