- A2A (Agent2Agent)
- An open protocol for communication between independent AI agents: discovering capabilities, delegating tasks and returning results. It complements MCP, which connects agents to tools. Details are version-dependent.Learn more: A2A (Agent2Agent) Protocol · A2A vs MCPSee also: MCP (Model Context Protocol)
- Agent (AI agent)
- A system in which a model pursues a goal by repeatedly choosing actions, calling tools and reading the results until it finishes or hits a limit.See also: Tool · Agentic workflow
- Agentic workflow
- A designed process that combines model calls, tools, checks and sometimes humans along a path the developer defines, as opposed to a model freely choosing every step.See also: Agent
- AI (Artificial Intelligence)
- The broad field of building software that performs tasks normally associated with human judgement, such as perception, language and decision-making.Learn more: What is Artificial Intelligence?See also: Machine learning · Deep learning
- Alignment
- Work to make AI systems behave in line with human intentions and values, including being helpful, honest and safe. Training steps such as instruction tuning and preference tuning contribute to it.Learn more: What is a Large Language Model (LLM)?See also: Instruction tuning · RLHF
- API key
- A secret string a program sends to an API to identify itself. Anyone who has the key can call the API as that program, so it belongs on a server, not in a browser.Learn more: API Keys · Environment Variables
- Approximate nearest neighbour (ANN)
- A family of search methods that quickly find vectors that are very likely, though not guaranteed, to be the closest to a query. Used by vector databases to stay fast at scale.Learn more: What is a Vector Database?See also: Vector database
- Attention
- The mechanism in a transformer that lets each token weigh how relevant every other token is when building its representation.Learn more: What is a Transformer? (Attention Explained)See also: Transformer
- Authorization code flow
- The usual OAuth 2.0 flow for user login: the browser gets a short-lived code, and the server exchanges that code for tokens. Public clients add PKCE so a stolen code is not enough.Learn more: OAuth 2.0 · OpenID Connect
- Backpropagation
- The algorithm that computes how each parameter contributed to the loss, so gradient descent can update it.Learn more: Backpropagation and Gradient DescentSee also: Neural network
- Benchmark
- A standard test used to compare models on a defined task. Useful for rough comparison, but not a substitute for testing on your own use case.Learn more: How to Evaluate AI Systems and OutputsSee also: Evals
- Best-of-N sampling
- Generating several candidate answers and returning the one a scorer — tests, a verifier or a reward model — rates highest.Learn more: Best-of-N SamplingSee also: Self-consistency
- C# (C sharp)
- A statically typed language on .NET, commonly used for APIs, workers and Azure services in Microsoft-centred stacks.Learn more: C#
- C2PA (Content Credentials)
- An open standard for cryptographically signed provenance metadata describing how media was made and edited, including AI generation.See also: Diffusion model
- Chain of thought
- A prompting technique in which the model is encouraged to work through intermediate reasoning steps before giving its answer. Its usefulness varies by model and task.See also: Prompt
- Chatbot
- Software that converses with users. Modern chatbots are usually powered by LLMs; the word says little about whether the system can take actions.See also: Agent
- Chunking
- Splitting documents into smaller passages so each can be embedded and retrieved on its own, as in RAG.
- Constitutional AI
- A training approach in which a model critiques and revises its outputs against written principles, with AI-generated preference labels (RLAIF) used for further training.Learn more: Constitutional AI and RLAIF (AI Feedback)See also: RLHF
- Constrained decoding (guided generation)
- Restricting a model's next-token choices at each step so output must follow a schema, regex or grammar. It guarantees well-formed output, not correct content.See also: Structured output
- Container
- A process isolated with its own filesystem view, built from an image and run by a container runtime. It shares the host kernel, unlike a virtual machine.Learn more: Containers · Docker
- Context engineering
- Curating everything a model sees each turn — instructions, retrieved facts, memory, tool definitions and results — so agents stay reliable.Learn more: Context Engineering · What is a Context Window?See also: Context window
- Context window (context length)
- The maximum amount of text, measured in tokens, a model can consider in one request, covering instructions, conversation, documents and the reply.Learn more: What is a Context Window? · What are Tokens in AI?See also: Token
- Contrastive learning
- Training that pulls matching pairs (such as an image and its caption) together in embedding space and pushes non-matching pairs apart. CLIP is a well-known example.Learn more: Contrastive Learning and CLIPSee also: Embedding
- CORS (Cross-Origin Resource Sharing)
- A browser check that asks a server whether a page from another origin may read the response. It is not authentication, and non-browser clients ignore it.Learn more: CORS
- Cosine similarity
- A measure of how closely two vectors point in the same direction. It is widely used to compare embeddings: values near 1 mean very similar, near 0 mean unrelated.Learn more: What are Embeddings in AI?
- Deep learning
- Machine learning that uses neural networks with many layers. It underlies most modern vision, speech and language models.Learn more: What is Artificial Intelligence?See also: Machine learning · Neural network
- Diffusion model
- A generative model, common in image generation, that starts from random noise and refines it step by step into content matching a prompt.Learn more: What is Generative AI?See also: Generative AI
- Direct Preference Optimization (DPO) (DPO)
- A preference-optimization method that tunes a model directly on chosen-versus-rejected response pairs relative to a reference model, without a separate reward model or RL loop. It is an alignment technique, not ordinary reinforcement learning.Learn more: Direct Preference Optimization (DPO) · Preference Optimization (Aligning Models from Human or AI Preferences)See also: RLHF
- Distillation (knowledge distillation)
- Training a smaller student model to imitate a larger teacher model's outputs, yielding a cheaper model for narrower tasks.Learn more: Knowledge DistillationSee also: Quantisation
- Embedding
- A list of numbers (a vector) that represents the meaning of content so that similar items are close together. Basis of semantic search and RAG.Learn more: What are Embeddings in AI? · What is a Vector Database?See also: Vector · Semantic search
- Embodied AI
- AI that perceives and acts in the physical or simulated world through a body such as a robot arm, drone or vehicle.Learn more: Embodied AISee also: VLA (vision-language-action model)
- Entra ID (Microsoft Entra ID, Azure AD)
- Microsoft's cloud identity service, formerly Azure Active Directory. It signs users in and issues tokens for Microsoft 365, Azure and registered apps.Learn more: Microsoft Entra ID · OAuth 2.0
- EU AI Act
- The European Union's risk-based regulation of AI systems, Regulation (EU) 2024/1689. Application dates are phased; verify the current timetable.Learn more: EU AI ActSee also: Guardrails
- Evals (evaluations)
- Structured tests that measure how well an AI system performs on tasks you care about, using test cases, graders and tracked scores.Learn more: How to Evaluate AI Systems and OutputsSee also: Benchmark
- Express
- A small Node.js library for HTTP routes and middleware. It does not include a database or a front end.
- Few-shot prompting
- Including a small number of example inputs and outputs in the prompt to show the model the pattern you want. Zero-shot means giving no examples.See also: Zero-shot prompting · Prompt
- Fine-tuning
- Further training a pre-trained model on your own examples so it adopts a consistent behaviour, format or skill.
- FlashAttention
- An exact attention algorithm that tiles computation to cut memory traffic, speeding up transformers and enabling longer sequences without approximating the result.Learn more: FlashAttentionSee also: Attention
- Foundation model
- A large model trained on broad data that can be adapted to many tasks. Large language models are the best-known example.Learn more: What is a Large Language Model (LLM)? · What is Generative AI?See also: LLM
- Function calling (tool calling)
- A model-API feature in which the model returns a structured request to run a named function with arguments. Your code runs it and returns the result.
- Generative AI
- AI that produces new content such as text, images, audio, video or code.Learn more: What is Generative AI?See also: LLM · Diffusion model
- GGUF
- A model file format used by llama.cpp, supporting many quantisation levels and bundling metadata for local inference.Learn more: llama.cpp · Quantization of AI ModelsSee also: Quantisation
- Git
- A distributed version-control system that records snapshots of a project so you can branch, merge and recover history.
- Graph neural network (GNN)
- A neural network that learns on graph-structured data by passing messages between connected nodes, used for molecules, networks and some weather models.Learn more: Graph Neural Networks (GNNs)See also: Neural network
- GraphQL
- An API style in which the client names the fields it wants and the server resolves them against a schema. Not the same thing as Microsoft Graph.Learn more: GraphQL · REST versus GraphQL
- GraphRAG
- Retrieval that builds a graph of entities and relationships (often with community summaries) so questions needing connections or corpus-wide synthesis can be answered.Learn more: GraphRAG (Knowledge-Graph Retrieval)See also: RAG (Retrieval-Augmented Generation)
- Grounding
- Tying a model's answer to supplied source material or tool results so claims can be traced and checked.See also: RAG · Hallucination
- GRPO (Group Relative Policy Optimization)
- A PPO-style reinforcement-learning method that normalises rewards within a group of sampled outputs per prompt, avoiding a separate value model. Used for training on tasks with verifiable rewards.See also: PPO (Proximal Policy Optimization)
- Guardrails
- Checks and limits around an AI system, such as input filters, output validation, permission limits and human approvals, that reduce harmful or unwanted behaviour.See also: Prompt injection
- Hallucination
- Fluent, confident model output that is false, unsupported by the sources, or invented, such as fake citations or facts.See also: Grounding
- Hugging Face (HF)
- A hub and set of libraries for publishing and loading machine-learning models and datasets, including the transformers library.Learn more: Hugging Face for Builders
- Human in the loop
- A design in which a person reviews or approves AI output or actions at defined points, especially before high-impact or irreversible steps.See also: Agentic workflow
- Hybrid search
- Combining keyword search with vector similarity (often plus metadata filters), so exact terms and meaning are both covered.See also: Semantic search
- Inference
- Running a trained model to produce output for new input, as opposed to training it.Learn more: What is a Large Language Model (LLM)? · What is Local AI? Running Models on Your Own DeviceSee also: Training
- Instruction tuning
- Additional training that teaches a pre-trained model to follow instructions and respond helpfully, turning a text predictor into an assistant.See also: Fine-tuning · RLHF
- Jailbreak
- An attempt to get a model to break its safety rules through crafted input. Related to, but distinct from, prompt injection.Learn more: What is Prompt Injection?See also: Prompt injection
- JSON Web Token (JWT)
- A signed token whose payload is JSON. The payload is readable by anyone; trust comes only from checking the signature, issuer, audience and expiry.Learn more: JSON Web Tokens · OAuth 2.0
- Knowledge cut-off
- The point after which a model's training data contains no information. Later events are unknown to it unless supplied via the prompt, search or tools.Learn more: What is a Large Language Model (LLM)?See also: LLM
- KV cache
- Stored attention keys and values for earlier tokens, reused during generation to avoid recomputation. It speeds decoding but grows with context length and batch size.Learn more: KV Cache · Prompt CachingSee also: Context window
- LangChain
- A framework for composing model calls, tools and retrieval in Python or JavaScript. It is optional glue, not a model.Learn more: LangChain, Practically · RAG Frameworks
- llama.cpp
- A C/C++ inference engine for running quantised models on CPUs and consumer GPUs using the GGUF format.Learn more: llama.cpp · GBNF Grammars (llama.cpp)See also: GGUF
- LlamaIndex (Llama Index)
- A framework focused on loading documents, indexing them and retrieving context for a language model.Learn more: LlamaIndex for Retrieval · RAG Frameworks
- LLM (Large Language Model)
- A very large neural network trained on text to predict the next token, enabling writing, summarising, translation, question answering and code generation.See also: Token · Transformer
- LLM observability
- Recording prompts, retrieved context, tool calls, outputs, latency and cost to debug and improve AI applications, while protecting sensitive data in those records.See also: Evals
- LLM-as-a-judge
- Using a language model to grade or compare other outputs against a rubric. It scales evaluation but has biases and needs calibration against human judgement.Learn more: LLM-as-a-JudgeSee also: Evals
- Local AI
- Running models on hardware you control instead of a hosted service, so data need not leave your device.Learn more: What is Local AI? Running Models on Your Own Device · Local AI vs Cloud AI: Privacy, Cost, CapabilitySee also: Open weights · Quantisation
- LoRA (low-rank adaptation)
- A parameter-efficient fine-tuning method that trains small add-on matrices while the original model stays frozen, greatly reducing training cost.Learn more: What is Fine-Tuning an AI Model?See also: Fine-tuning
- Machine learning
- The part of AI in which systems learn patterns from data rather than following hand-written rules.Learn more: What is Artificial Intelligence?See also: Deep learning · Training
- MCP (Model Context Protocol)
- An open protocol for connecting AI applications to external tools, data and prompts through servers that clients can discover and call.
- MCP client
- The component inside an AI host application that holds a connection to one MCP server and relays requests and results.See also: MCP · MCP server
- MCP server
- A program that exposes tools, resources and prompts to MCP clients, often wrapping an existing API, database or file system.See also: MCP · MCP client
- Mechanistic interpretability
- Research that tries to explain model behaviour by analysing internal components such as features and circuits.Learn more: Mechanistic InterpretabilitySee also: Neural network
- Memory (agent)
- Information a system stores outside the model and re-inserts into the prompt, such as recent messages, summaries and durable facts, because models themselves retain nothing between requests.Learn more: Agent Memory: Short-Term, Long-Term and BeyondSee also: Context window
- Microsoft 365 (M365, Office 365)
- Microsoft's cloud suite for mail, files, SharePoint and Teams. Office 365 is the older name for the same tenant story.Learn more: Microsoft 365 for Developers · Microsoft Graph API
- Microsoft Graph
- The REST API at graph.microsoft.com for Microsoft 365 data such as users, mail, files, sites and Teams. It is not GraphQL.Learn more: Microsoft Graph API · Microsoft 365 for Developers
- Mixture of experts (MoE)
- A model design with many specialist sub-networks and a router that activates only a few per token, so total parameters exceed active parameters.Learn more: Mixture of Experts (MoE)See also: Transformer
- MMLU
- A multiple-choice benchmark across 57 subjects used to compare breadth of knowledge in language models; now close to saturated for top models.See also: Benchmark
- Model card
- A structured document describing a model's intended use, training data, evaluation, limitations and ethical considerations.Learn more: Model Cards and Datasheets (AI Documentation)See also: Open weights
- Model drift (data drift, concept drift)
- Degradation of a deployed model's performance as input data or the relationship between inputs and outcomes changes over time.Learn more: Model Drift and Production MonitoringSee also: Evals
- Multi-agent system
- A setup in which several model-driven agents with distinct roles or contexts cooperate on a larger task.Learn more: Multi-Agent Systems ExplainedSee also: Agent
- Multimodal
- Able to handle more than one type of data, such as text together with images or audio.Learn more: What is Multimodal AI?See also: Generative AI
- Neural network
- A model made of layers of simple connected units whose strengths (parameters) are adjusted during training. The basis of deep learning.See also: Deep learning · Parameters
- Next.js
- A React framework that adds routing and a server runtime so secrets and data fetching can stay off the browser.
- OAuth 2.0 (OAuth)
- A protocol for granting an application limited access to an API without sharing the user's password. OpenID Connect adds sign-in identity on top.
- Open weights
- Models whose trained parameters are published for download and local use, subject to the licence. Not always the same as fully open source.See also: Local AI
- OpenID Connect (OIDC)
- An identity layer on OAuth 2.0. The ID token says who signed in. The access token is what you send to an API.Learn more: OpenID Connect · OAuth 2.0
- ORM (object-relational mapper)
- A library that maps database rows to objects in a programming language. Prisma is one TypeScript example. The SQL still runs.Learn more: ORMs and Prisma · SQL
- ORM (outcome reward model)
- A scorer that judges only the final answer of a reasoning attempt. Not to be confused with an object-relational mapper in software.See also: PRM (process reward model)
- Overfitting
- When a model or prompt works well on the examples used to build it but poorly on new cases. Held-out test data helps detect it.See also: Evals
- Parameters (weights)
- The numbers inside a model that are adjusted during training and determine its behaviour. Model size is often quoted as a parameter count.Learn more: What is a Large Language Model (LLM)?See also: Neural network · Training
- Perplexity
- A language-modelling metric reflecting how well a model predicts text; lower means less "surprised". It is not a measure of usefulness or truth.See also: Token
- Positional encoding
- Information added so a transformer knows token order; methods include sinusoidal encodings, RoPE and ALiBi.Learn more: Positional Encoding (Sinusoidal, RoPE, ALiBi)See also: Transformer
- PostgreSQL (Postgres, PG)
- An open-source relational database. Applications talk to it with SQL. The pgvector extension can store embeddings in the same server.
- Power Platform
- Microsoft's low-code family: Power Apps, Power Automate, Power BI and Dataverse, used for tenant workflows beside Microsoft 365.Learn more: Power Platform Basics
- PPO (Proximal Policy Optimization)
- A widely used reinforcement-learning algorithm that clips policy updates to keep training stable; used in RLHF pipelines.Learn more: Proximal Policy Optimization (PPO)See also: RLHF
- Pre-training
- The initial, large-scale training stage in which a model learns general patterns from broad data before any specialised adaptation.See also: Fine-tuning · Training
- Preference optimization
- The family of methods that tune models using comparisons of better versus worse responses — RLHF, DPO, KTO, IPO, ORPO, SimPO and relatives.See also: Direct Preference Optimization (DPO)
- PRM (process reward model)
- A scorer that rates each intermediate step of a reasoning chain, enabling step-level feedback and guided search.See also: ORM (outcome reward model)
- Prompt
- The input given to a model: instructions, context, examples and the user's request.Learn more: Prompt Engineering: How to Write Better AI Prompts · Common Prompting Mistakes (and How to Fix Them)See also: System prompt
- Prompt caching
- Reusing computation for a repeated prompt prefix to cut latency and often cost. Rules and pricing are provider-specific.Learn more: Prompt CachingSee also: KV cache
- Prompt injection
- An attack where text supplied to a model, directly or hidden in content it reads, tries to override the developer's instructions.Learn more: What is Prompt Injection?See also: Guardrails · Jailbreak
- Quantisation
- Storing a model's parameters at lower numeric precision to cut memory use, usually at some cost in quality.See also: Local AI · Parameters
- RAG (Retrieval-Augmented Generation)
- Retrieving relevant passages from your data and giving them to a model so its answer is based on that material.
- Reasoning model
- A language model trained or configured to work through a problem in intermediate steps before answering, trading latency and cost for accuracy on multi-step tasks.Learn more: Reasoning Models · Reasoning Models vs Standard LLMsSee also: Chain of thought
- Red teaming
- Structured adversarial testing of an AI system to find failures, misuse paths and vulnerabilities before real attackers or users do.Learn more: Red-Teaming AI SystemsSee also: Prompt injection
- Redis
- An in-memory store for keys, expirations, counters and short-lived coordination. Best used as a cache or rate limiter beside a durable database.Learn more: Redis
- Reinforcement learning (RL)
- Learning to choose actions to maximise cumulative reward through trial and feedback. Preference methods such as DPO are classified separately as preference optimization.Learn more: Reinforcement LearningSee also: PPO (Proximal Policy Optimization)
- Reranking
- A second, more precise scoring step that reorders retrieved candidates before they are given to the model.Learn more: What is RAG (Retrieval-Augmented Generation)?See also: RAG
- Reward hacking
- When a system exploits flaws in its reward signal, scoring highly without doing what was intended.Learn more: Reward Hacking and Over-OptimizationSee also: RLHF
- RLAIF (reinforcement learning from AI feedback)
- Training with preference labels generated by an AI model, guided by written principles, instead of human labels.Learn more: Constitutional AI and RLAIF (AI Feedback)See also: RLHF
- RLHF (reinforcement learning from human feedback)
- A training approach that uses human preference judgements to steer a model towards responses people rate as better.See also: Instruction tuning · Alignment
- RLVR (reinforcement learning from verifiable rewards)
- Reinforcement learning where rewards come from automatic checkers such as exact answer match or unit tests, used to train reasoning behaviour.
- Sampling
- The step where the next token is chosen from the model's probabilities. Settings such as temperature control how predictable or varied the choice is.Learn more: What is a Large Language Model (LLM)?See also: Temperature · Token
- Scaling laws
- Empirical relationships showing how model loss tends to improve predictably with parameters, data and compute.Learn more: Scaling LawsSee also: Pre-training
- Self-consistency
- Sampling several reasoning attempts and choosing the final answer most of them agree on.Learn more: Self-Consistency (Sampling and Voting)See also: Best-of-N sampling
- Semantic search
- Searching by meaning rather than exact words, usually by comparing embeddings.Learn more: What are Embeddings in AI? · What is a Vector Database?See also: Embedding · Hybrid search
- Small language model (SLM)
- A compact language model, typically up to a few billion parameters, that runs cheaply and can be strong on narrow tasks.Learn more: Small Language Models (SLMs)See also: LLM (Large Language Model)
- Speculative decoding
- Speeding up generation by letting a small draft model propose tokens that a large model verifies in parallel, preserving the large model's output distribution.Learn more: Speculative DecodingSee also: Inference
- SPFx (SharePoint Framework)
- The short name for the SharePoint Framework, used to build client-side SharePoint web parts and extensions with TypeScript, usually with React.Learn more: SharePoint Framework (SPFx) · How to Build an SPFx Web Part
- State space model (SSM)
- A sequence model that carries a compact recurrent state, giving linear-time processing as an alternative to attention; Mamba is a well-known example.See also: Attention
- Structured output
- Model output constrained to a defined format, such as JSON matching a schema, so software can use it directly.Learn more: Function Calling (Tool Calling) Explained · Prompt Engineering: How to Write Better AI PromptsSee also: Function calling
- Sycophancy
- A model's tendency to tell users what they want to hear — agreeing with mistaken claims or flattering them — rather than what is accurate.Learn more: Sycophancy in Language ModelsSee also: RLHF
- System prompt
- Standing instructions from the developer that define a model's role, rules and format across a conversation, separate from the user's messages.Learn more: What is a System Prompt?See also: Prompt
- Temperature
- A setting that controls randomness when a model picks the next token. Lower values give more predictable output; higher values give more varied output.See also: Sampling
- Test-time compute (inference-time scaling)
- Spending extra computation when answering — longer reasoning, multiple samples, search or verification — to improve hard answers.See also: Reasoning model
- Thinking budget (reasoning effort)
- A provider setting that caps or tunes how much reasoning a model does per request. Names and semantics vary by provider.Learn more: Thinking Budgets and Reasoning EffortSee also: Reasoning model
- Token
- The unit of text a model reads and writes: a word, part of a word, or punctuation. Limits and prices are commonly measured in tokens.Learn more: What are Tokens in AI?See also: Tokenizer · Context window
- Tokenizer
- The component that splits text into tokens and maps them to numbers. Different model families use different tokenizers.Learn more: What are Tokens in AI?See also: Token
- Tool (agent tool)
- A described capability, such as search, a database query or an API call, that a model can request and that your code executes.See also: Function calling · MCP
- Training
- The process of adjusting a model's parameters using data so its outputs improve on a target task.See also: Inference · Machine learning
- Transformer
- The neural-network architecture, introduced in 2017, behind most modern language models. Its key feature is attention.Learn more: What is a Transformer? (Attention Explained)
- TypeScript (TS)
- JavaScript with a static type system. The compiler emits plain JavaScript. SharePoint Framework projects are TypeScript-first.Learn more: TypeScript · SharePoint Framework (SPFx)
- Vector
- An ordered list of numbers. In AI, embeddings are vectors, and closeness between vectors reflects similarity in meaning.Learn more: What are Embeddings in AI?See also: Embedding · Cosine similarity
- Vector database (vector store)
- A system that stores embedding vectors with metadata and finds the closest matches to a query vector quickly.See also: Embedding · Approximate nearest neighbour (ANN)
- Vision Transformer (ViT)
- A transformer that splits an image into patches treated as tokens, matching or beating CNNs at scale.Learn more: Vision Transformers (ViT) · CNN vs Vision TransformerSee also: Transformer
- Vision-language model (VLM)
- A model that takes images (and often video) with text and produces text: descriptions, answers about charts and screenshots, visual reasoning.Learn more: Vision-Language Models (VLMs)See also: Multimodal
- VLA (vision-language-action model)
- A model that maps camera images and a language instruction to robot actions, extending vision-language models to control.Learn more: Vision-Language-Action (VLA) ModelsSee also: Embodied AI
- vLLM
- An open-source LLM serving engine known for paged KV-cache memory and continuous batching.Learn more: vLLM · Local and Self-Hosted LLM Runtimes ComparedSee also: KV cache
- Webhook
- An HTTP callback: another system POSTs to your URL when an event happens. Verify the signature and treat retries as normal.Learn more: Webhooks
- World model
- A learned model of how an environment evolves in response to actions, used for planning, imagination-based training or generated simulation.Learn more: World ModelsSee also: Embodied AI
- Zero-shot prompting
- Asking a model to perform a task with instructions only and no examples.See also: Few-shot prompting