[{"data":1,"prerenderedAt":220},["ShallowReactive",2],{"glossary":3,"ill-generic\u002Fglossary":219},[4,12,19,26,32,37,44,49,54,59,65,70,75,80,86,91,96,101,106,111,116,121,127,133,138,143,148,153,159,164,169,174,179,185,190,195,200,205,209,214],{"term":5,"slug":6,"definition":7,"related":8},"RAG (Retrieval-Augmented Generation)","rag","RAG is a technique where a language model retrieves relevant passages from your own documents before generating an answer, so responses are grounded in current, private data and can cite sources. It reduces hallucinations and avoids retraining models each time content changes. It is the standard approach for company knowledge assistants and enterprise search.",[9,10,11],"rag-systems","generative-ai-llm","ai-chatbots",{"term":13,"slug":14,"definition":15,"related":16},"LLM (Large Language Model)","llm","A large language model is a neural network trained on vast amounts of text to understand and generate language. Models such as GPT, Claude, Gemini and Llama power chat assistants, summarisation, extraction and code generation. Businesses use LLMs through APIs or self-hosted open-source versions, usually combined with retrieval, tools and guardrails.",[10,17,18],"ai-integration","open-source-ai",{"term":20,"slug":21,"definition":22,"related":23},"Agentic AI","agentic-ai","Agentic AI describes systems in which a language model plans, decides and takes actions through tools (calling APIs, reading data, sending messages) to complete multi-step goals rather than answering a single prompt. Agents run loops of reasoning and action with human checkpoints where needed, and are used for research, operations, support and sales workflows.",[24,25,10],"ai-agents","ai-automation",{"term":27,"slug":28,"definition":29,"related":30},"Multi-agent system","multi-agent-system","A multi-agent system splits work between several specialised AI agents (for example a planner, a researcher, a writer and a reviewer) coordinated by an orchestrator. Each agent has its own instructions and tools. This improves reliability on complex tasks and makes behaviour easier to test and monitor than one large prompt.",[24,31],"mlops-ai-infrastructure",{"term":33,"slug":34,"definition":35,"related":36},"Vector database","vector-database","A vector database stores embeddings (numeric representations of text, images or audio) and finds the items most similar to a query in milliseconds. It is the retrieval layer behind RAG, semantic search and recommendation systems. Options include Pinecone, Weaviate, Qdrant, Milvus, ChromaDB and PostgreSQL with pgvector.",[9,17],{"term":38,"slug":39,"definition":40,"related":41},"Embeddings","embeddings","Embeddings are dense numeric vectors that capture the meaning of text, images or other data so that similar items sit close together in vector space. They are produced by embedding models and power semantic search, clustering, recommendations and RAG retrieval. Custom embeddings can be trained for specialised domains.",[9,42,43],"nlp","ai-fine-tuning",{"term":45,"slug":46,"definition":47,"related":48},"Hybrid search","hybrid-search","Hybrid search combines keyword (lexical) search with vector (semantic) search and merges the results, often with a re-ranking model. It captures exact matches such as product codes and names as well as meaning-based matches, giving better relevance than either method alone. Most production RAG systems use hybrid retrieval.",[9,42],{"term":50,"slug":51,"definition":52,"related":53},"Fine-tuning","fine-tuning","Fine-tuning continues training a pre-trained model on your own examples so it learns a domain, style, format or task more reliably than prompting alone. It is used for consistent output formats, specialised terminology, smaller and cheaper models, and classification or extraction tasks. It complements, rather than replaces, retrieval.",[43,18],{"term":55,"slug":56,"definition":57,"related":58},"LoRA (Low-Rank Adaptation)","lora","LoRA is an efficient fine-tuning method that trains small adapter matrices instead of all model weights, cutting GPU memory and cost dramatically while keeping most of the quality of full fine-tuning. Adapters can be swapped per task or client, making LoRA the usual choice for customising open-source LLMs and image models.",[43,18],{"term":60,"slug":61,"definition":62,"related":63},"RLHF (Reinforcement Learning from Human Feedback)","rlhf","RLHF trains a model to prefer outputs that humans rate as better by learning a reward model from human comparisons and optimising against it. It is how assistant models become helpful and safe. Businesses apply lighter variants, such as preference tuning, to align custom models with their quality standards.",[43,64],"ai-consulting-strategy",{"term":66,"slug":67,"definition":68,"related":69},"Quantisation","quantisation","Quantisation reduces the numeric precision of model weights, for example from 16-bit to 4-bit, so a model needs less memory and runs faster with a small loss in accuracy. It allows large open-source models to run on cheaper GPUs, on-premise servers or edge devices, and lowers inference cost.",[18,31],{"term":71,"slug":72,"definition":73,"related":74},"Inference","inference","Inference is the act of running a trained model to produce an output: a prediction, a classification or generated text. In production, inference cost, latency and throughput determine how an AI feature performs and what it costs. Optimising inference involves batching, caching, quantisation and the right hardware.",[31,18],{"term":76,"slug":77,"definition":78,"related":79},"MLOps","mlops","MLOps is the set of practices and tools for deploying, monitoring and maintaining machine-learning and AI systems in production: versioning data and models, automated testing and deployment, monitoring quality and drift, and managing infrastructure. It turns a working prototype into a reliable, maintainable service.",[31,17],{"term":81,"slug":82,"definition":83,"related":84},"Model drift","model-drift","Model drift is the decline in a model’s accuracy over time as real-world data changes: new customer behaviour, products, seasons or fraud tactics. Monitoring compares live inputs and outputs against training-time baselines and triggers retraining or review when drift is detected. It is a core MLOps concern.",[31,85],"data-analytics-predictive-ai",{"term":87,"slug":88,"definition":89,"related":90},"Hallucination","hallucination","A hallucination is a confident but false or unsupported statement generated by a language model. It happens because models predict plausible text rather than verify facts. Mitigations include retrieval with citations, restricting answers to trusted sources, guardrails, evaluation sets and human review for high-stakes outputs.",[9,10,31],{"term":92,"slug":93,"definition":94,"related":95},"Prompt engineering","prompt-engineering","Prompt engineering is the practice of designing instructions, examples and context given to a language model so it produces reliable, well-formatted outputs. It covers system prompts, few-shot examples, output schemas and prompt templates, and is tested with evaluation sets. It is the first lever before fine-tuning.",[10,24],{"term":97,"slug":98,"definition":99,"related":100},"Tool use \u002F function calling","tool-use-function-calling","Tool use, or function calling, lets a language model request that specific functions be run (search a database, call an API, send an email) and use the results in its response. It is the mechanism that turns chat models into agents that can act inside business systems, with permissions and logging.",[24,17,25],{"term":102,"slug":103,"definition":104,"related":105},"Guardrails","guardrails","Guardrails are checks placed around an AI system to keep behaviour within agreed limits: filtering inputs and outputs, blocking off-topic or unsafe content, enforcing formats, validating facts against sources and requiring human approval for sensitive actions. They are essential for customer-facing and regulated deployments.",[11,24,31],{"term":107,"slug":108,"definition":109,"related":110},"Evaluation set","evaluation-set","An evaluation set is a curated collection of test inputs and expected outputs used to measure an AI system’s quality objectively. It is run before release and after every change to catch regressions, and combines automated scoring with human review. Building one from real cases is a first step in serious AI delivery.",[31,10,64],{"term":112,"slug":113,"definition":114,"related":115},"Token","token","A token is the unit of text a language model reads and writes, roughly three-quarters of an English word. Model pricing, context limits and latency are measured in tokens, so prompt design, retrieval size and output length directly affect cost. Token counts also drive rate limits on hosted APIs.",[10,17],{"term":117,"slug":118,"definition":119,"related":120},"Context window","context-window","The context window is the maximum number of tokens a model can consider at once, including instructions, retrieved documents, conversation history and its own reply. Larger windows allow more material but cost more and can dilute attention, so RAG and summarisation are used to keep context focused.",[10,9],{"term":122,"slug":123,"definition":124,"related":125},"Speech-to-text (STT)","speech-to-text","Speech-to-text, or automatic speech recognition, converts spoken audio into written text in real time or from recordings. It underpins voice agents, call transcription, meeting notes and speech analytics. Modern models handle accents, noise and many languages, and can be run through APIs or self-hosted.",[126,42],"voice-ai",{"term":128,"slug":129,"definition":130,"related":131},"TTS (Text-to-Speech)","tts","Text-to-speech converts written text into natural-sounding spoken audio. Combined with speech-to-text and a language model it completes a voice agent that can hold conversations on the phone or in apps. Voices can be selected or cloned, and latency and pronunciation are tuned per use case.",[126,132],"multimodal-ai",{"term":134,"slug":135,"definition":136,"related":137},"IVR (Interactive Voice Response)","ivr","IVR is the phone menu system that greets callers and routes them by keypad or spoken input. AI IVR replaces rigid menus with a conversational voice agent that understands natural requests, answers common questions and routes intelligently, reducing hold times and misroutes while keeping human transfer available.",[126,11],{"term":139,"slug":140,"definition":141,"related":142},"NER (Named Entity Recognition)","ner","Named entity recognition identifies and labels entities in text such as people, organisations, locations, dates, amounts and product names. It is used to structure documents, extract fields from contracts and invoices, redact personal data and enrich search. LLMs and specialised NLP models are both used, depending on volume and precision needs.",[42,25],{"term":144,"slug":145,"definition":146,"related":147},"Sentiment analysis","sentiment-analysis","Sentiment analysis classifies text or speech as positive, negative or neutral, and often detects specific emotions or aspects, for example delivery versus product quality. It is applied to reviews, support tickets, calls and social media to track customer experience and prioritise responses.",[42,85,11],{"term":149,"slug":150,"definition":151,"related":152},"Intent detection","intent-detection","Intent detection determines what a user wants from a message (check an order, cancel a booking, ask for pricing) so a chatbot, voice agent or routing system can respond correctly. It is implemented with classifiers or LLM prompting and evaluated on real conversations to reduce misroutes.",[42,11,126],{"term":154,"slug":155,"definition":156,"related":157},"OCR (Optical Character Recognition)","ocr","OCR converts images of printed or handwritten text (scans, photos, PDFs) into machine-readable text. Modern OCR combined with layout understanding and LLMs powers document processing for invoices, forms, IDs and contracts, extracting structured fields rather than raw text alone.",[158,42,25],"computer-vision",{"term":160,"slug":161,"definition":162,"related":163},"Object detection","object-detection","Object detection locates and classifies objects in images or video, drawing bounding boxes around items such as products, vehicles, people or defects. It powers quality inspection, shelf analytics, CCTV analytics and licence-plate recognition, and can run in real time on edge devices.",[158,18],{"term":165,"slug":166,"definition":167,"related":168},"Image segmentation","image-segmentation","Image segmentation labels every pixel in an image by class or object, producing precise outlines rather than boxes. It is used for medical imaging, defect measurement, agricultural and drone analysis and background removal, wherever exact shape and area matter.",[158,132],{"term":170,"slug":171,"definition":172,"related":173},"Pose estimation","pose-estimation","Pose estimation detects the position of body joints in images or video to understand human posture and movement. Applications include workplace safety monitoring, sports and physiotherapy analysis, retail behaviour analytics and interactive experiences. It runs on standard cameras without wearable sensors.",[158,132],{"term":175,"slug":176,"definition":177,"related":178},"Time-series forecasting","time-series-forecasting","Time-series forecasting predicts future values from historical sequences (demand, load, revenue, traffic) using statistical and machine-learning models that learn trend, seasonality and the effect of external drivers. Forecasts are delivered with confidence ranges and monitored for accuracy as new data arrives.",[85,31],{"term":180,"slug":181,"definition":182,"related":183},"Churn prediction","churn-prediction","Churn prediction estimates which customers are likely to cancel or stop buying, based on usage, support, billing and engagement signals. Scores and their drivers are delivered to customer-success and marketing teams so retention effort goes to the accounts where it changes the outcome.",[85,184],"ai-saas-product-development",{"term":186,"slug":187,"definition":188,"related":189},"Anomaly detection","anomaly-detection","Anomaly detection identifies data points or patterns that differ significantly from normal behaviour: unusual transactions, sensor readings, log events or meter data. It works with or without labelled examples and underpins fraud detection, predictive maintenance and monitoring, with alerts prioritised by severity.",[85,31],{"term":191,"slug":192,"definition":193,"related":194},"Knowledge graph","knowledge-graph","A knowledge graph represents entities (customers, products, documents, people) and the relationships between them as a network. Combined with RAG, it lets AI systems answer questions that require connecting facts across documents and reason over structure, improving precision in complex enterprise domains.",[9,42],{"term":196,"slug":197,"definition":198,"related":199},"On-prem AI","on-prem-ai","On-premise AI runs models on your own servers or private cloud rather than through a third-party API, so sensitive data never leaves your environment. It relies on open-source models, GPU infrastructure and MLOps practices, and suits healthcare, legal, finance, defence and any organisation with strict data-residency requirements.",[18,31],{"term":201,"slug":202,"definition":203,"related":204},"Edge AI","edge-ai","Edge AI runs models directly on devices such as cameras, industrial PCs, phones or gateways instead of in the cloud. It gives millisecond latency, works without connectivity and keeps data local. Models are optimised through quantisation and pruning to fit device constraints, and results sync centrally when possible.",[18,158,31],{"term":206,"slug":132,"definition":207,"related":208},"Multimodal AI","Multimodal AI processes and generates more than one type of data (text, images, audio and video) in a single system. Examples include describing images, answering questions about documents with charts, generating video from scripts and voice assistants that see and hear. It broadens what AI can automate.",[132,158,126],{"term":210,"slug":211,"definition":212,"related":213},"AI governance","ai-governance","AI governance is the set of policies, roles, controls and records that ensure AI systems are used responsibly and in line with regulation: model inventories, risk assessments, data-handling rules, human oversight, monitoring and audit trails. It gives leadership confidence to scale AI without unmanaged risk.",[64,31],{"term":215,"slug":216,"definition":217,"related":218},"Human-in-the-loop","human-in-the-loop","Human-in-the-loop design keeps people involved at key points in an AI workflow (reviewing outputs, approving actions, labelling uncertain cases) so errors are caught and the system keeps learning. It is the default pattern for high-stakes decisions and for building trust before increasing automation.",[25,24,64],"\u003Csvg xmlns=\"http:\u002F\u002Fwww.w3.org\u002F2000\u002Fsvg\" viewBox=\"0 0 640 480\" role=\"img\" aria-hidden=\"true\">\u003Cg fill=\"none\" stroke=\"currentColor\" stroke-width=\"5\" stroke-linecap=\"round\" stroke-linejoin=\"round\">\u003Cpath d=\"M120 130Q220 105.8 320 130V350Q220 325.8 120 350Z\"\u002F>\u003Cpath d=\"M320 130Q420 105.8 520 130V350Q420 325.8 320 350Z\"\u002F>\u003Cg stroke-width=\"3.5\">\u003Cline x1=\"150\" y1=\"166\" x2=\"290\" y2=\"166\" stroke-opacity=\"0.45\"\u002F>\u003Cline x1=\"150\" y1=\"194\" x2=\"262\" y2=\"194\" stroke-opacity=\"0.45\"\u002F>\u003Cline x1=\"150\" y1=\"222\" x2=\"283\" y2=\"222\" stroke-opacity=\"0.45\"\u002F>\u003Cline x1=\"150\" y1=\"250\" x2=\"248\" y2=\"250\" stroke-opacity=\"0.45\"\u002F>\u003Cline x1=\"150\" y1=\"278\" x2=\"276\" y2=\"278\" stroke-opacity=\"0.45\"\u002F>\u003Cline x1=\"350\" y1=\"166\" x2=\"490\" y2=\"166\" stroke-opacity=\"0.45\"\u002F>\u003Cline x1=\"350\" y1=\"194\" x2=\"476\" y2=\"194\" stroke-opacity=\"0.45\"\u002F>\u003Cline x1=\"350\" y1=\"222\" x2=\"455\" y2=\"222\" stroke-opacity=\"0.45\"\u002F>\u003Cline x1=\"350\" y1=\"250\" x2=\"483\" y2=\"250\" stroke-opacity=\"0.45\"\u002F>\u003Cline x1=\"350\" y1=\"278\" x2=\"441\" y2=\"278\" stroke-opacity=\"0.45\"\u002F>\u003C\u002Fg>\u003Cg stroke-width=\"3.5\">\u003Crect x=\"520\" y=\"150\" width=\"24\" height=\"32\" rx=\"6\"\u002F>\u003Crect x=\"520\" y=\"194\" width=\"24\" height=\"32\" rx=\"6\"\u002F>\u003Crect x=\"520\" y=\"238\" width=\"24\" height=\"32\" rx=\"6\"\u002F>\u003Crect x=\"520\" y=\"282\" width=\"24\" height=\"32\" rx=\"6\"\u002F>\u003C\u002Fg>\u003Ccircle cx=\"532\" cy=\"210\" r=\"6\" fill=\"var(--accent, #0E7490)\" stroke=\"none\"\u002F>\u003Ccircle cx=\"490\" cy=\"340\" r=\"44\"\u002F>\u003Cline x1=\"521.1\" y1=\"371.1\" x2=\"560.7\" y2=\"410.7\" stroke-width=\"11\"\u002F>\u003Cpath d=\"M464.4 330.7A27.3 27.3 0 0 1 485.3 313.1\" stroke-opacity=\"0.45\"\u002F>\u003C\u002Fg>\u003C\u002Fsvg>",1792973682537]