AI glossary: OCR, IDP, RAG, agentic AI and more

Plain-language definitions of the AI terms you will hear when buying or building AI for your business. Each entry links to the related service or article.

Jump to: A B C D E F G H I L M N O P R S T V W

Agentic AI

Agentic AI describes systems that do more than answer a question. They plan a few steps, call tools such as your CRM or a search, check their own output and carry on until a task is finished. The more freedom the system has, the more important limits, logging and approval steps become.

See how we build these inside business process automation.

API

An API (application programming interface) is a defined way for one piece of software to ask another for data or an action. Most AI projects depend on APIs to read from and write to your ERP, CRM, accounting package or WhatsApp Business account.

Integration work is part of every custom software build.

Business intelligence (BI)

Business intelligence is the practice of turning raw business data into charts, tables and alerts that managers can act on. A BI dashboard might show sales by branch, defects by shift or pending invoices by age. It works best when the underlying data is clean and refreshed automatically.

Read about our BI dashboards and custom software.

Computer vision

Computer vision is the branch of AI that lets software understand images and video. It is used to inspect parts, count items, read gauges, check safety gear and monitor stock from camera feeds. It needs good images as much as a good model.

Explore our computer vision services.

Confidence score

A confidence score is the number a model attaches to its own answer, showing how sure it is. Low-confidence results are usually sent to a person to check, while high-confidence ones pass straight through. Setting that threshold is a business decision, not just a technical one.

Used in every document AI workflow we build.

Data labelling

Data labelling means marking examples so a model can learn from them, such as drawing boxes around defects or tagging fields on an invoice. Good labels, applied consistently, matter more than a large pile of messy ones.

Part of how we prototype computer vision systems.

Data pipeline

A data pipeline is the automated route that data takes from its source to where it is used, with cleaning and checks along the way. Pipelines keep dashboards and models fed without someone exporting spreadsheets every week.

Built as part of our BI and software work.

DPDP Act

The Digital Personal Data Protection Act, 2023 is India's law on handling personal data. It affects how AI systems collect, store and process customer and employee information. Confirm the exact obligations for your business with your legal adviser.

See our approach in the build vs buy guide and our privacy policy.

Edge AI

Edge AI means running a model on a device near where the data is created, such as a camera box on a factory line or a phone, instead of in the cloud. It gives fast responses, works without a steady internet connection and keeps raw data local.

Used in computer vision and in Kepton, which runs on the phone.

Embeddings

Embeddings are lists of numbers that capture the meaning of a piece of text or an image. Two items with similar meaning end up with similar numbers, so software can search by meaning instead of exact keywords. They power most document search and RAG systems.

Core to our generative AI assistants.

Fine-tuning

Fine-tuning means training an existing model a little further on your own examples so it behaves in a specific way, such as using your tone or classifying your document types. It is often unnecessary: good prompts and RAG solve many problems for less effort and cost.

We weigh the options in generative AI projects.

Generative AI

Generative AI creates new content, such as text, images, summaries or code, based on patterns learned from large amounts of data. For business use it works best when grounded in your own documents and checked by a person before anything reaches a customer.

Explore generative AI and custom assistants.

GST reconciliation

GST reconciliation is the matching of your purchase invoices with what suppliers have filed, so you can claim input tax credit correctly. It is a classic task for document AI, because the data starts on paper or PDF and must match exactly.

Handled with document AI and automation.

Hallucination

A hallucination is a confident answer from a language model that is wrong or invented. It happens because the model predicts likely text and does not look facts up. Grounding answers in your documents, showing sources and keeping a person in the loop reduces the risk.

We design against it in generative AI builds.

Human-in-the-loop

Human-in-the-loop means a person reviews or approves the AI's work at chosen points, usually when confidence is low or the stakes are high. It makes a system safer at launch and gives you labelled corrections that improve it over time.

A design principle across document AI and inspection projects.

IDP (intelligent document processing)

Intelligent document processing goes beyond reading text. It classifies each document, extracts the right fields, validates them against rules and your master data, and sends clean records into your systems. Exceptions go to a person for review.

Read OCR vs IDP or see our document AI service.

Inference

Inference is the moment a trained model is used to produce an answer, as opposed to training, when it learns. Inference cost and speed decide whether a system is affordable and responsive at scale.

Considered in every Webonika service scope.

LLM (large language model)

A large language model is an AI system trained on very large amounts of text to understand and write language. It can summarise, draft, translate and answer questions. On its own it does not know your business, which is why it is usually combined with your documents through RAG.

See generative AI and custom assistants.

Machine learning

Machine learning is the approach of letting software learn patterns from examples instead of following hand-written rules. Most practical AI today, from defect detection to demand forecasting, is machine learning applied to a specific business problem.

Applied across our AI services.

Model drift

Model drift is the slow loss of accuracy that happens when real-world data changes after a model is trained, for example new packaging, new invoice layouts or seasonal behaviour. Monitoring and periodic retraining keep it in check.

This is why we offer an annual maintenance contract; see pricing.

NLP (natural language processing)

Natural language processing covers software that reads, interprets and produces human language. Classifying support tickets, extracting names from emails and understanding a customer's message in Hinglish are all NLP tasks.

Used in voice agents and assistants.

Object detection

Object detection finds and locates specific things in an image or video, usually by drawing a box around each one. It is used to count parts, spot a missing helmet, find a defect on a surface or check what is on a shelf.

Core to computer vision inspection and safety projects.

OCR (optical character recognition)

OCR converts the text in a scanned page or photograph into text a computer can read. It tells you what the characters are, but not what they mean, so it cannot tell an invoice number from a PO number on its own.

Compare it with IDP in OCR vs IDP, or see document AI.

Prompt

A prompt is the instruction and context you give a language model. Clear prompts, with examples and constraints, change the quality of answers a great deal. In a business system the prompt is written once, tested and kept under version control.

Part of every generative AI build.

RAG (retrieval-augmented generation)

RAG lets a language model answer using your own documents. The system first retrieves the most relevant passages, then asks the model to write an answer based only on them and to cite the source. It keeps answers current and reduces hallucination without retraining the model.

This is the core of our generative AI assistants.

RPA (robotic process automation)

RPA uses software robots to repeat the clicks and keystrokes a person makes on a screen, such as copying data between systems. It is good for stable, rule-based steps. AI is added when the input is messy, such as a scanned form or a free-text email.

See business process automation.

STT (speech-to-text)

Speech-to-text turns spoken audio into written text. Accuracy depends on accent, language, background noise and call quality, so it must be tested on real recordings of your own callers, including Indian languages and mixed Hindi-English speech.

The listening half of our voice AI agents.

Synthetic data

Synthetic data is artificially created data used to train or test a model when real examples are rare or sensitive. It can help, for instance, with unusual defect types, but it supplements real data and does not replace it.

Sometimes used in computer vision prototypes.

Token

A token is the small chunk of text, often part of a word, that a language model reads and writes. Usage is measured and priced in tokens, so long documents and long conversations cost more. Hindi and other Indian languages can use more tokens per sentence than English.

We estimate this when we cost generative AI projects.

TTS (text-to-speech)

Text-to-speech turns written text into spoken audio. In a voice agent it is how the system talks back. Natural pacing, correct pronunciation of names and good Indian-language voices matter more than raw audio quality.

The speaking half of our voice AI agents.

Vector database

A vector database stores embeddings and finds the closest matches quickly. It is the memory that lets a RAG assistant find the right paragraph among thousands of pages in a fraction of a second.

Used in document assistants.

Webhook

A webhook is an automatic message one system sends to another when something happens, such as a new lead arriving or a payment succeeding. Webhooks are how automations start without anyone polling or pressing a button.

Used in automation workflows.