Short definitions of the AI, software, marketing and operations terms that appear across this site. Each links to the service or article where it matters.
An AI agent is software that uses a large language model to decide which steps to take and which tools to call, such as searching a database, drafting a reply or updating a record, to complete a task. Unlike a chatbot, it acts as well as answers. Production agents need guardrails, logging and a clear path to a human when they are unsure.
Human in the loop is a design where a person reviews, approves or corrects an AI system's output before it takes effect. It is used where a wrong action would be costly, for example sending money, changing a customer record or replying on a regulated topic.
A large language model is a neural network trained on large amounts of text to predict the next word, which lets it write, summarise, classify and answer questions. Claude, GPT, Gemini, Llama and Mistral are examples. On its own, a model knows only what was in its training data, which is why business systems connect it to their own documents and tools.
Retrieval-augmented generation is a pattern where a system first retrieves relevant passages from your own documents, then gives them to a language model to write the answer. It lets the model answer from current, private information and cite its sources. When a RAG system gives a confident wrong answer, retrieval is the usual cause, not the model.
Chunking is how documents are split into pieces before they are indexed for retrieval. Splitting on the document's own structure, such as sections, clauses or table rows, keeps related facts together. Splitting at a fixed size can separate a table from its header and make retrieval fail.
An embedding is a list of numbers that represents the meaning of a piece of text, so that texts with similar meanings sit close together. Retrieval systems compare the embedding of a question with the embeddings of stored chunks to find relevant passages.
A vector database stores embeddings and quickly finds the ones nearest to a query. Pinecone, Weaviate, Qdrant and pgvector, an extension for PostgreSQL, are common choices.
A hallucination is when a language model states something untrue or unsupported by its sources, with full confidence. It is reduced, not removed, by grounding answers in retrieved documents, requiring citations, and testing against known questions before each release.
An evaluation harness is a set of automated tests that scores an AI system's answers against expected results before a change ships. A good harness runs on every release, including prompt edits, and blocks the release if quality drops.
A golden set is a curated list of real questions with checked, correct answers, used as the benchmark inside an evaluation harness. It is reviewed and extended as new failure cases appear.
Guardrails are rules and checks around an AI system that limit what it can say or do: blocking unsafe topics, redacting personal data, requiring citations, or stopping an agent from taking an action without approval.
Prompt injection is an attack where text in a user message, web page or document tries to override an AI system's instructions, for example telling an agent to ignore its rules or reveal data. Defences include separating instructions from content, limiting which tools an agent may call, and testing with known attack examples.
A model router sends each request to the language model best suited to it, weighing quality, speed and cost. Simple requests can go to a cheaper, faster model and hard ones to a stronger model, without changing the application.
A token is the unit a language model reads and writes, roughly three quarters of an English word on average. Model prices, speed and limits are all counted in tokens.
The context window is the amount of text a language model can take in at once, measured in tokens. A larger window lets the model read more retrieved passages, but retrieval still decides which passages it sees.
Fine-tuning is further training of an existing model on your own examples so it adopts a style, a format or a narrow skill. It changes how a model behaves; it is not the best way to give it new facts, which retrieval does better.
PII redaction is removing or masking personally identifiable information, such as names, phone numbers, ID numbers and addresses, before data is stored, logged or sent to a model.
A minimum viable product is the smallest version of a product that lets a team test its riskiest assumption with real users. It is a learning tool, not a cut-down final product, and it is judged by what it teaches the team.
Technical debt is the future cost of shortcuts taken in code or architecture to ship sooner. Like financial debt, it can be a sensible trade, but left unpaid it slows every later change and raises the risk of defects.
An application programming interface (API) is a defined way for one piece of software to request data or actions from another. Integrations between a website, a CRM and a payment provider, for example, run through APIs.
A cross-platform app is built from one codebase, with a framework such as React Native or Flutter, and released on both iOS and Android. It usually costs less to build and maintain than two native apps, at some cost in access to the newest platform features.
A progressive web app is a website that can be installed on a phone's home screen and keep working on a poor connection or offline, using standard web technology instead of an app store release.
A design system is a shared library of reusable interface components, with the rules for using them, that keeps a product consistent as more people build it. It usually covers colour, type, spacing, components and accessibility guidance.
Web accessibility means people with disabilities can perceive, navigate and use a website, including with a screen reader, a keyboard alone or magnification. The Web Content Accessibility Guidelines (WCAG), published by the W3C, are the usual standard.
Continuous integration and delivery is the practice of merging code changes often, testing each change automatically, and keeping the main branch ready to release, so that deployments are small, routine and easy to roll back.
A data warehouse is a central database built for analysis. It holds cleaned data from many source systems, so that reports, dashboards and models all read from one place. Snowflake and BigQuery are common examples.
ELT moves raw data from source systems into a warehouse first and transforms it there, often with version-controlled SQL in a tool such as dbt. ETL, the older pattern, transforms data before loading it.
Core Web Vitals are Google's three measures of page experience: Largest Contentful Paint (LCP), how quickly the main content appears; Interaction to Next Paint (INP), how quickly the page responds to input; and Cumulative Layout Shift (CLS), how much the layout jumps. Google rates a page good when LCP is within 2.5 seconds, INP within 200 milliseconds and CLS 0.1 or less, at the 75th percentile of visits.
Technical SEO is the work that lets search engines crawl, render and index a site correctly: clean URLs and redirects, canonical tags, sitemaps, structured data, fast pages, and content that is present in the HTML rather than loaded only by JavaScript.
Generative engine optimisation is making a website easy for AI answer engines, such as Google's AI Overviews, ChatGPT search, Perplexity and Copilot, to find, understand, trust and cite. It builds on SEO and adds clear, consistent facts, answer-first writing, structured data and open access for the crawlers these engines use.
Structured data is machine-readable labelling, usually JSON-LD using the schema.org vocabulary, that tells search engines what a page describes: an organisation, a service, an article or a set of questions and answers. It must match what the page visibly says.
E-E-A-T stands for experience, expertise, authoritativeness and trustworthiness: the qualities Google's search quality rater guidelines ask raters to judge in a page and its creator. Google describes trust as the most important of the four.
Conversion rate optimisation is increasing the share of visitors who take a chosen action, such as booking a call or buying, by testing changes to pages, offers and forms against the current version.
Marketing attribution assigns credit for a lead or sale to the channels and campaigns that came before it. Models range from last click to data-driven; each answers a slightly different question, and none is the whole truth on its own.
Customer acquisition cost is the total sales and marketing spend in a period divided by the number of new customers won in that period. It is most useful read alongside how much a customer is worth over time.
Return on ad spend is the revenue attributed to advertising divided by what the advertising cost. A ROAS of 4 means four units of revenue for every unit spent. Blended ROAS divides all revenue by all ad spend, which avoids arguments about attribution.
A service level agreement is a written commitment to measurable service levels, such as response time, resolution time or availability, and to what happens if they are missed.
Customer satisfaction score is the result of a short survey after an interaction, usually asking customers to rate their satisfaction on a scale such as 1 to 5. It is reported as the share of satisfied responses or as an average.
First response time is how long a customer waits for the first useful reply after contacting support. It is measured per channel, because expectations differ between chat, email and phone.
A network operations centre is a team, often working around the clock, that monitors systems and networks, responds to alerts and escalates incidents before they reach users.
Infrastructure as code means defining servers, networks and cloud services in version-controlled configuration files, using tools such as Terraform, instead of setting them up by hand. Changes can be reviewed, repeated and rolled back like application code.
Business process outsourcing is handing a business function, such as customer support, data entry or back-office processing, to an outside team that runs it to agreed service levels.
Tell us where you are, where you want to go, and the deadlines you cannot miss. We'll respond within one business day with a clear next step.
Direct line
support@telematrixglobal.com
+91 79808 07674
Office hours
Mon to Sat · 09:00 to 19:00 IST
Support runs 24/7.
