Saraswati Stitch®contact@saraswatistitch.comThe Resource · Glossary
AI, in plain words.
The terms clients ask us to translate, defined the way we would say them out loud. No jargon for its own sake, and a note on where each one actually earns its place.
Agentic AI
Agentic AI refers to systems where a model plans and carries out a multi-step task on its own, deciding and acting in a loop, rather than producing a single response to a single prompt.
AI agent
An AI agent is a system that uses a language model to decide which steps to take and which tools to call to finish a goal, instead of just answering a single prompt.
AI copilot
An AI copilot is an assistant embedded in a tool or workflow that suggests, drafts, and helps in real time while a human stays in control and makes the final call.
API
An API is a defined way for two pieces of software to talk to each other, and it is how an AI system reads from and writes to the tools a business already runs.
Artificial general intelligence (AGI)
Artificial general intelligence is a hypothetical AI that can understand and perform any intellectual task a human can, as opposed to today’s AI, which is narrow and task-specific.
Chain of thought
Chain of thought is prompting a model to work through a problem step by step before giving its answer, which improves accuracy on anything that needs reasoning.
Computer vision
Computer vision is the field of AI that lets machines interpret images and video, from reading a document to recognising an object or inspecting a product for defects.
Context window
The context window is the maximum amount of text, measured in tokens, that a model can consider at once, covering the instructions, the conversation, and any documents you provide.
Deep learning
Deep learning is machine learning that uses many-layered neural networks to learn complex patterns, and it is the approach behind most modern AI breakthroughs.
Diffusion model
A diffusion model is the kind of AI that generates images, and increasingly video, by starting from noise and refining it step by step into a result that matches a prompt.
Embeddings
An embedding is a list of numbers that captures the meaning of a piece of text, so that passages about similar things sit close together and can be found by similarity rather than exact words.
Evaluation (evals)
An evaluation, or eval, is a repeatable test that measures how well an AI system does its job on real examples, so you can tell whether a change made it better or worse.
Few-shot prompting
Few-shot prompting is giving a model a handful of worked examples inside the prompt so it follows the pattern, without any retraining.
Fine-tuning
Fine-tuning is further training a base model on your own examples so it learns a specific style, format, or task more reliably than prompting alone can achieve.
Function calling
Function calling is when a model, instead of answering in text, returns a structured request to run one of the tools you have given it, which is how an AI takes real actions.
Generative AI
Generative AI is artificial intelligence that creates new content, such as text, images, audio, or video, rather than only classifying or predicting from existing data.
Grounding
Grounding is tying a model’s answer to a verifiable source, such as your documents or live data, so it reports facts instead of inventing them.
Guardrails
Guardrails are the rules and checks that keep an AI system inside safe, correct behaviour, blocking off-topic, unsafe, or out-of-scope responses before they reach the user.
Hallucination
A hallucination is when a model states something false or invented as if it were fact, because it predicts plausible-sounding text rather than looking anything up.
Human in the loop
Human in the loop is a design where an AI system does the work but a person reviews, approves, or handles the cases the system is unsure about, keeping judgement and accountability with a human.
Inference
Inference is the act of running a trained model to get an output, and inference cost is what you pay each time the model answers, measured in tokens.
Knowledge base
A knowledge base is the curated collection of documents an AI system draws on to answer questions, and keeping it accurate and current is what keeps the answers accurate and current.
Knowledge graph
A knowledge graph stores facts as connected entities and relationships, letting a system answer questions by following links between things rather than matching text.
Large language model (LLM)
A large language model is a neural network trained on huge amounts of text to predict the next token, which lets it write, summarise, classify, and answer questions across almost any topic.
Latency
Latency is how long a model takes to produce its answer, and in live uses like chat or voice it matters as much as the quality of the answer itself.
Machine learning
Machine learning is a branch of AI where a system learns patterns from data and improves at a task through examples, instead of being programmed with explicit rules.
Model Context Protocol (MCP)
The Model Context Protocol is an open standard that lets an AI model connect to external tools and data sources through one consistent interface, instead of a custom integration for each one.
Model distillation
Distillation trains a smaller, cheaper model to mimic a larger one on a specific task, so you keep most of the quality at a fraction of the cost and latency.
Model training
Model training is the process of teaching an AI model by showing it data and adjusting its internal parameters until it performs the task well, which is done once up front and is computationally expensive.
Multimodal AI
Multimodal AI is a model that can take in or produce more than one kind of data, such as text plus images, audio, or video, rather than text alone.
Natural language processing (NLP)
Natural language processing is the field of AI concerned with letting computers understand, interpret, and generate human language.
Neural network
A neural network is a computing system loosely inspired by the brain, made of layers of connected nodes that adjust their strengths as they learn from data.
OCR (optical character recognition)
OCR turns an image of text, such as a scanned invoice or a photo of a form, into machine-readable text, which is the first step in automating any paper-based process.
Open-source model
An open-source model is one whose weights are published so you can run it on your own infrastructure, giving you control over data and cost at the price of hosting it yourself.
Orchestration
Orchestration is the layer that coordinates several models, tools, and steps into one working pipeline, deciding what runs when and passing results from one stage to the next.
Prompt engineering
Prompt engineering is the craft of writing the instructions, examples, and context you give a model so it produces reliable, correctly formatted output for a specific task.
Prompt injection
Prompt injection is an attack where hidden or malicious instructions in content the model reads trick it into ignoring its real instructions, which is a core security risk for any AI that reads untrusted data.
Quantization
Quantization shrinks a model by storing its numbers at lower precision, so it runs faster and cheaper on smaller hardware with a small, usually acceptable, drop in quality.
Rate limit
A rate limit is the cap a provider puts on how many requests or tokens you can use in a window, and designing around it is part of running any AI system reliably at scale.
Reranking
Reranking is a second pass that reorders search or retrieval results by how well they actually answer the query, lifting the quality of what a RAG system shows the model.
Retrieval-augmented generation (RAG)
Retrieval-augmented generation is a technique that fetches relevant passages from your own documents at question time and hands them to the model, so its answer is grounded in your data instead of its training.
RLHF (reinforcement learning from human feedback)
RLHF is a training method where people rank a model’s outputs and the model is tuned toward the preferred ones, which is a big part of why modern assistants are helpful and well-behaved.
Semantic search
Semantic search finds results by meaning rather than exact keywords, so a search for "cancel my plan" surfaces the "how to end your subscription" page even though they share no words.
Speech to text
Speech to text converts spoken audio into written text, which is what lets a voice agent understand a caller and what turns calls and meetings into searchable records.
Structured output
Structured output is when a model is made to return data in a strict format such as JSON, so another system can use the result directly instead of parsing free text.
Synthetic data
Synthetic data is artificially generated examples used to train or test a model when real data is scarce, sensitive, or imbalanced.
System prompt
A system prompt is the standing instruction given to a model before the conversation starts, setting its role, rules, tone, and the boundaries it must keep across every reply.
Temperature
Temperature is a setting that controls how random a model’s output is: low temperature gives consistent, predictable answers, high temperature gives more varied, creative ones.
Token
A token is the unit a language model reads and writes in, roughly a word-piece of a few characters, and model pricing and limits are counted in tokens rather than words.
Transformer
A transformer is the neural-network design behind modern language models, able to weigh the relationships between all parts of an input at once, which is what made large language models possible.
Vector database
A vector database stores embeddings and finds the closest matches to a query vector quickly, which is what lets a RAG system retrieve the right passages from millions of documents in milliseconds.
Voice agent
A voice agent is an AI system that answers or makes phone calls, understands speech, holds a natural conversation, and takes actions like booking or rescheduling, usually handing off to a human when needed.
Workflow automation
Workflow automation is the practice of moving a repeatable business process from manual steps to a system that runs it end to end, with a person involved only where judgement is actually needed.
Zero-shot learning
Zero-shot learning is when a model performs a task it was not given any examples for, relying on its general training to figure it out from the instruction alone.