Reference · AI & Software
AI Development Glossary: Key AI & Software Terms
Plain-English definitions of the terms used in AI and software development projects — RAG, LLMs, embeddings, AI agents, vector databases, prompt injection, multi-tenant SaaS, and more.
This glossary is written for business and technical decision-makers evaluating AI and software projects. Each entry gives a concise definition and, where useful, why it matters in practice. It is a living reference: terms are grouped by topic so you can jump to the area you need.
In one line: an AI development glossary explains the vocabulary behind AI systems — how models are built and called, how they access your data, how they act through agents, and how they are secured and scaled.
Core AI
AI and Machine Learning Terms
Artificial Intelligence (AI)
Software that performs tasks normally requiring human intelligence, such as understanding language, recognising patterns, or making recommendations. In business, AI is usually applied to one specific task rather than as general reasoning.
Machine Learning (ML)
A branch of AI where a model learns patterns from data instead of being explicitly programmed. ML is used for forecasting, classification, anomaly detection, and recommendation.
Deep Learning
Machine learning using multi-layered neural networks. It underpins most modern language, vision, and speech systems.
Model
The trained artefact that turns input into output — for example, a model that classifies a support ticket or predicts next month’s demand. A model is not the same as the application around it.
Training vs Inference
Training is the process of teaching a model from data. Inference is using the trained model to produce an answer. Inference cost and latency are usually the figures that matter most in production.
Fine-tuning
Further training of an existing model on your own examples to adapt its behaviour or style. It is often less necessary than good prompt design and retrieval, and it carries ongoing maintenance cost.
Generative AI
Large Language Models and Generative AI
Large Language Model (LLM)
A model trained on large amounts of text to understand and generate language. Examples include GPT-4o, Claude, and Gemini. LLMs are accessed through APIs or run on your own infrastructure.
Generative AI
AI that produces new content — text, summaries, answers, code, or images. In business software it commonly powers assistants, drafting tools, and summarisation.
Prompt
The instruction or input given to a model. Prompt quality strongly affects output quality, which is why prompt design is treated as an engineering activity, not an afterthought.
Prompt Engineering
The practice of structuring instructions, examples, and constraints so a model produces reliable results. It includes system prompts, output formats, and refusal behaviour.
Context Window
The maximum amount of text a model can consider at once, measured in tokens. It limits how much source material can be included in a single request and influences cost.
Token
A unit of text — roughly part of a word — that models read and bill against. Token volume is a primary driver of LLM running cost.
Hallucination
When a model states something incorrect with confidence. Grounding answers in retrieved sources, requiring citations, and allowing refusal are the main mitigations.
Guardrails
Controls that constrain model behaviour, such as allowed topics, output validation, and escalation to a human. Guardrails make AI usable in customer-facing and regulated contexts.
Retrieval
RAG and Retrieval Terms
Retrieval-Augmented Generation (RAG)
An approach that retrieves relevant passages from your own data and supplies them to an LLM so answers are grounded in real sources. RAG is usually the fastest route to a reliable organisational assistant.
Embedding
A numeric representation of text (or other data) that captures meaning, allowing similar items to be found by similarity rather than exact keywords.
Vector Database
A database optimised to store embeddings and perform fast similarity search. Options include Pinecone and PostgreSQL with pgvector; the right choice depends on scale, cost, and existing infrastructure.
Semantic Search
Search based on meaning rather than keyword matching. It finds relevant results even when the wording differs, which is why it suits knowledge assistants.
Chunking
Splitting documents into smaller passages before embedding. Chunk size and overlap significantly affect retrieval quality and should be tuned, not defaulted.
Reranking
A second pass that reorders retrieved results by relevance before they reach the model. Reranking often improves answer accuracy more than changing the model.
Agents & automation
AI Agents and Automation Terms
AI Agent
A system that pursues a goal by deciding on and executing multiple steps, often using tools and external systems. Agents can read data, take actions, and update records, so access control matters.
Tool Use / Function Calling
The ability of a model to call defined functions or APIs — for example, to look up a record or send a message. This is how agents interact with real systems safely and predictably.
Multi-Agent System
A design where several specialised agents cooperate on a workflow. Useful for complex processes, but it should be justified by the problem, not adopted by default.
Workflow Automation
Connecting systems so a process runs with minimal manual effort — for example, an enquiry that creates a CRM record, sends a reply, and notifies a team. Automation is where AI often shows the clearest return.
Software & cloud
Software, API, and Cloud Terms
API (Application Programming Interface)
A defined way for systems to communicate. APIs are how AI features connect to your CRM, ERP, databases, and internal tools.
REST and GraphQL
Two common API styles. REST is widely supported and simple to cache; GraphQL lets clients request exactly the data they need. The choice depends on the consumers of the API.
Multi-Tenant SaaS
A software product where one application serves many customers while keeping their data logically separate. Tenant isolation, billing, and permissions are core architectural concerns.
Microservices
Building an application as a set of independently deployable services. It improves scaling and team autonomy but increases operational and security surface.
CI/CD and DevOps
Continuous integration and delivery automate testing and deployment; DevOps is the practice around it. Together they enable frequent, low-risk releases with rollback options.
Scaling
Increasing capacity to handle load. Stateless services, caching, and queue-based processing allow applications to scale horizontally as usage grows.
Security & privacy
AI Security and Privacy Terms
Prompt Injection
An attack where malicious input tries to override a model’s instructions. Defences include input and output filtering, scoped tool access, and separating trusted instructions from user content.
PII (Personally Identifiable Information)
Data that can identify an individual. Handling PII responsibly means collecting only what is needed, controlling access, and limiting retention.
Data Minimisation
The principle of using only the data a feature genuinely requires. It reduces both privacy risk and the cost of storing and processing data.
Access Control / RBAC
Restricting what users and services can see or do, often by role. In AI systems it also governs which data a model can retrieve and which tools an agent may call.
FAQ
Glossary Questions
What is the difference between AI and machine learning?
AI is the broad goal of software performing intelligent tasks. Machine learning is one method of achieving it, where a model learns patterns from data rather than following explicit rules. Most practical business AI today is machine learning.
What is RAG in simple terms?
RAG, or retrieval-augmented generation, finds relevant passages in your own documents and gives them to a language model so its answer is grounded in real sources. It reduces hallucination and enables citations.
Do I need a vector database for AI?
Not always. You need one when you are doing semantic search or RAG at a scale where a general database is too slow. PostgreSQL with pgvector is sufficient for many projects; dedicated vector databases help at larger scale.
What is an AI agent, and is it different from a chatbot?
A chatbot answers questions. An AI agent also takes actions — looking up records, calling tools, and updating systems to complete a multi-step task. Agents require stricter access control because they can change data.
What is fine-tuning, and when is it worth it?
Fine-tuning further trains a model on your own examples. It is worth considering for consistent style or specialised tasks, but good prompting and retrieval solve most needs with less cost and maintenance.
What is prompt injection?
An attack where crafted input tries to override a model’s instructions. It is mitigated with input and output filtering, scoped tool permissions, and keeping trusted instructions separate from user-supplied content.
Putting terms to work
How We Apply These Concepts
Glossaries only help if the concepts translate into working systems. Innovative AI Labs builds RAG assistants, AI agents, automation, and full-stack software using the terms above — with grounding, evaluation, access control, and monitoring designed in.
Explore our services: AI automation and RAG development, custom software and SaaS development, mobile apps, and cloud and DevOps. See how we deliver for teams in the USA, UK, and Australia. Planning a budget? Try our AI cost estimator.
Need help turning these concepts into a working product — a RAG assistant, an AI agent, or an automated workflow? Talk to our team about your requirements and technical challenges.