AI glossary

A

Agentic AI

Agentic AI describes AI systems that pursue a goal over several steps, choosing which tools to call and what to do next, instead of answering a single prompt.

AI agent

An AI agent is a system where a language model decides which actions to take, calls tools to take them, looks at the results and repeats until a goal is reached or it hands over to a person.

AI governance

AI governance is the set of policies, roles, records and controls an organisation uses to decide which AI it uses, for what, with which data, and who is accountable when it goes wrong.

AI orchestration

AI orchestration is the layer that coordinates models, tools, data and people in a multi-step AI process: what runs in which order, what each step receives, and what happens when a step fails.

C

Computer vision

Computer vision is the field of AI that lets software interpret images and video: find objects, count them, read them, measure them and flag what looks wrong.

Context window

The context window is the maximum amount of text, counted in tokens, that a language model can take in and produce in a single request, including instructions, documents, chat history and its own answer.

D

Diffusion model

A diffusion model is a generative AI model that creates images, video or audio by starting from random noise and removing it step by step until an output matching the prompt appears.

DPDP Act 2023

The Digital Personal Data Protection Act, 2023 is India's data protection law: it sets how businesses may collect and use personal data in digital form, based on notice and consent, with penalties set by a Data Protection Board.

E

E-invoicing in India (IRN)

E-invoicing under GST means reporting each B2B invoice to a government Invoice Registration Portal, which returns an Invoice Reference Number (IRN) and a signed QR code that must appear on the invoice.

Edge AI

Edge AI means running AI models on devices close to where data is created, such as cameras, gateways and small boards, instead of sending everything to a cloud server.

Embeddings

An embedding is a list of numbers that represents the meaning of a piece of text, an image or audio, so that similar items end up close together and a computer can compare them by distance.

F

Fine-tuning

Fine-tuning is further training of an already trained model on a smaller set of your own examples so it learns a specific task, format or style.

Function calling (tool use)

Function calling is a model feature where the language model, given descriptions of available functions, returns a structured request to call one with specific arguments, which your code then runs.

G

GPU

A GPU (graphics processing unit) is a processor with thousands of small cores that run the same maths in parallel, which makes it the standard hardware for training and running AI models.

GSTR-1 vs GSTR-3B

GSTR-1 is the GST return that lists your outward supplies invoice by invoice; GSTR-3B is the summary return in which you declare tax payable, claim input tax credit and pay the tax.

GSTR-2B reconciliation

GSTR-2B reconciliation is matching the purchase invoices in your books against GSTR-2B, the monthly input tax credit statement on the GST portal, so you claim only credit your suppliers have reported.

Guardrails (AI)

AI guardrails are checks placed around a model's inputs, outputs and actions that block, correct or escalate anything outside the rules a business has set.

H

Hallucination (AI)

An AI hallucination is output from a model that sounds confident and plausible but is false or unsupported, such as an invented citation, figure or policy.

Human in the loop

Human in the loop (HITL) is a design where an AI system proposes or prepares work and a person reviews, corrects or approves it before it takes effect.

I

Inference

Inference is running a trained model on new input to get an output, such as generating a reply, classifying an email or detecting objects in an image.

Intelligent document processing (IDP)

Intelligent document processing combines OCR, classification, AI extraction, validation and human review to turn documents such as invoices, forms and IDs into checked, structured data in your systems.

L

Large language model (LLM)

A large language model is a neural network trained on very large amounts of text to predict the next token, which lets it read, write, summarise and reason over language.

LoRA (low-rank adaptation)

LoRA is a fine-tuning method that freezes a model's original weights and trains a small set of extra low-rank matrices, so a model can be adapted on modest hardware and the adapter stored as a small file.

M

Mixture of experts

Mixture of experts (MoE) is a model design where each layer holds many specialist sub-networks and a router sends each token to only a few of them, so a very large model uses only a small part of itself per token.

Model Context Protocol (MCP)

The Model Context Protocol is an open standard that defines how AI applications connect to external tools and data sources through MCP servers, so one integration works with any compatible AI client.

Model distillation

Model distillation (knowledge distillation) trains a small "student" model to copy the outputs of a large "teacher" model, so you get much of the large model's skill at a fraction of the size and cost.

Model evaluation

Model evaluation is measuring how well an AI model or AI system performs on a fixed set of realistic test cases, so you can compare options and catch regressions before users do.

Multi-agent system

A multi-agent system is a setup where several AI agents, each with its own role, tools and permissions, work on parts of a task and pass requests and results between them.

O

OCR (optical character recognition)

OCR, optical character recognition, is technology that converts text in images and scanned documents into machine-readable text.

On-premise AI

On-premise AI means running AI models, including large language models, on hardware you own or control, so prompts, documents and outputs never pass through a third-party AI provider.

Open-weight model

An open-weight model is an AI model whose trained weights are published for download, so you can run, inspect and fine-tune it on your own hardware under the terms of its licence.

P

Prompt engineering

Prompt engineering is the practice of writing and testing the instructions, examples and context given to a language model so it produces the output a task needs, reliably.

Q

Quantization

Quantization stores a model's weights with fewer bits, for example 4-bit integers instead of 16-bit floats, so the model needs less memory and runs faster with a small loss in quality.

R

Retrieval-augmented generation (RAG)

Retrieval-augmented generation is a pattern where a system first searches your own documents for relevant passages and then gives them to a language model so its answer is grounded in that material.

Robotic process automation

Robotic process automation (RPA) is software that operates other applications through their screens, clicking, typing and copying like a person, to automate repetitive tasks where no API exists.

S

Semantic search

Semantic search finds results by meaning rather than exact words, so a search for "refund for a damaged parcel" also finds a document titled "returns policy for broken items".

Speech-to-text

Speech-to-text (automatic speech recognition, ASR) is AI that converts spoken audio into written text, often with timestamps and speaker labels.

Synthetic data

Synthetic data is artificially generated data that mimics the patterns of real data, used to train, test or demo AI systems when real data is scarce, sensitive or unbalanced.

T

TDS and TCS

TDS (tax deducted at source) is income tax a payer holds back from a payment and deposits for the payee; TCS (tax collected at source) is tax a seller adds to certain sales and deposits for the buyer.

Temperature

Temperature is a setting on a language model that controls how random its word choices are: low values give focused, repeatable answers, high values give more varied ones.

Text-to-speech

Text-to-speech (TTS) is AI that turns written text into natural sounding spoken audio, in a chosen voice and language.

Tokens

Tokens are the small pieces of text, often parts of words, that a language model reads and writes, and the unit in which context limits and API prices are measured.

V

Vector database

A vector database stores embeddings and finds the stored items closest to a query vector quickly, which is the search engine behind most semantic search and RAG systems.

VRAM

VRAM (video RAM) is the memory built into a GPU; an AI model's weights and working data must fit in it, so VRAM decides which models a machine can run.

W

Webhook

A webhook is an HTTP request one application sends to a URL you provide the moment something happens, such as a payment, an order or a new message, so your system can react instantly.

WhatsApp Business API

The WhatsApp Business API, now Meta's WhatsApp Business Platform, lets software send and receive WhatsApp messages on a business number, so CRMs, bots and several staff can share one number.

WhatsApp message template

A WhatsApp message template is a pre-approved message format, with placeholders for names, dates or amounts, that a business must use to message a customer outside the 24-hour customer service window.

Workflow automation

Workflow automation is software that runs a repeatable business process on its own: a trigger starts it, data moves between apps, rules decide the path, and people step in only where needed.