IonBin
All pages
Search
Directory

Concepts & glossary

Plain-English definitions with the business consequence of each term.

59 entriesReview-by target: 180 days

A

  • Affiliate and sponsorship disclosure

    An affiliate or sponsorship disclosure tells readers that you earn money or other value from a recommendation. The FTC's guidance says disclosures should be clear, close to the recommendation and hard to miss. A lone 'affiliate link' label or a single homepage notice may not be enough. IonBin's wiki pages carry no affiliate links at all.

  • Agent harness

    An agent harness is the software around a model that manages context, tools, working state and execution. Evaluate its permissions and checks alongside the model; a framework does not by itself establish reliability.

  • Agent skills (SKILL.md)

    An agent skill packages instructions and optional scripts or references in a folder with SKILL.md. Claude and Codex support the open format, but available tools, installation locations and vendor extensions differ. Test a shared skill in each intended client.

  • Agentic commerce

    Agentic commerce means an AI agent finds, compares and pays for something on a person's or company's behalf. Open protocols now exist for it: the Agentic Commerce Protocol (created by Stripe, OpenAI and Meta), Google's Universal Commerce Protocol, and x402 for paying for web content or APIs per request. Sellers need clear prices, scope and checkout that a machine can read.

  • AI agent

    An AI agent is a system in which a model chooses steps and tools toward a goal. Workflows use predefined orchestration, which can still include branches, checks and retries. Both need clear permissions, stopping conditions and evidence that the result is correct.

  • AI context, memory and orchestration

    Context supplies the information an assistant works from. Memory retains information across sessions. Orchestration coordinates tools, people and steps. Products and methods in this layer organize work around language models; they are not themselves new foundation models.

  • AI governance

    AI governance defines who may choose, use and monitor AI, with data rules, review requirements and accountable owners. A small business can start with a short policy. Voluntary frameworks and applicable law help determine what further controls are needed.

  • AI Overviews and AI Mode (Google)

    AI Overviews and AI Mode are Google Search features that generate answers with supporting links. Existing Search eligibility applies; Google does not require special AI markup. Their traffic is included in Search Console’s Web reporting.

  • Answer engine optimization (AEO / GEO)

    Answer engine optimization is the work of making your brand and content easy for AI answer engines to find, understand, trust and cite. It builds on SEO basics (crawlable, clear, helpful pages) and adds attention to what third-party sources say about you. Results are hard to measure precisely because AI answers vary.

  • API (application programming interface)

    An API defines how one software component requests data or actions from another. For business automation, check the exact operations, access requirements and limits; API availability alone does not prove a workflow can run.

B

  • Background agents

    A background agent performs a task without a person supervising each step. It may run in a cloud service or on a local or self-hosted machine, triggered by a schedule, event or explicit handoff. Check where it executes, what it can access and how its result is reviewed.

  • Batch processing (batch APIs)

    Batch APIs let you submit many AI requests at once and get the results later, in exchange for a lower price. OpenAI and Anthropic both charge 50% of standard prices for batch work, with a 24-hour window. Batches suit work that doesn't need an instant answer, such as tagging a backlog or generating descriptions overnight.

C

  • Computer use

    Computer use lets an AI system inspect a graphical interface and request clicks, typing and other actions. It can reach tools without a suitable API. Your application executes those actions, so permissions, isolation, outcome checks and confirmation for consequential actions are part of the implementation.

  • Content credentials (C2PA)

    Content credentials are tamper-evident metadata attached to an image, video or audio file that record where it came from and how it was edited, including whether AI was used. They follow the C2PA open standard. They help audiences judge authenticity, but metadata can be stripped, so they are a signal, not proof.

  • Context rot

    Context rot describes less reliable model performance as inputs grow, particularly when useful information competes with distractors. Research has demonstrated length and position effects in specific models and tasks. Relevant context and task-specific testing can help; there is no universal token threshold or guaranteed fix.

  • Context window

    The context window is how much material a model can consider in one request: your instructions, documents, the conversation so far and its own reply, measured in tokens. Current flagship API models from Anthropic and OpenAI list about 1M tokens. App plans can allow much less, so check the plan you use.

D

  • Data governance

    Data governance is how an organization decides who owns data, who can access it, how accurate it must be, and how long it is kept. AI raises the stakes: assistants and agents can read and copy data at machine speed. Good governance makes AI answers more trustworthy and leaks less likely.

  • Data provenance

    Data provenance records where information came from and how it changed: the original source, extraction or editing steps and review decisions. For a knowledge collection, retain useful page or timestamp references and permissions. A traceable source can still be wrong or outdated.

E

  • Embeddings

    An embedding represents text, an image or other data as numbers in a learned vector space. Distance or similarity measures can help find related items. Embeddings support semantic search, recommendations, clustering and RAG, but similarity does not establish truth or relevance for every task.

  • ETL and data pipelines

    ETL (extract, transform, load) is the process of pulling data from source systems, cleaning and reshaping it, and loading it somewhere useful, such as a spreadsheet or warehouse. It is the plumbing behind reliable reporting. AI can now handle messy transform steps, like categorizing free text.

  • Evals (AI evaluations)

    Evals are repeatable tests that check whether an AI system does a task well enough, using your own examples and a clear pass bar. They turn 'it seems fine' into evidence, and tell you whether a new model, prompt or vendor is actually better. A few dozen real examples is a practical starting point.

F

  • Fine-tuning

    Fine-tuning continues training an existing model for a task, behavior or domain. It can help with style, structure and narrow skills; frequently changing facts often fit retrieval better. Availability depends on the provider, model and account: OpenAI has announced a self-serve wind-down, and Mistral labels its legacy service deprecated.

G

  • Generative AI

    Generative AI is AI that creates new content, such as text, images, audio, video or code, in response to a prompt. Large language models are the best-known kind. For a business, it is a drafting and analysis engine, not a source of truth: outputs need checking.

  • Golden set: reviewed evaluation examples

    A golden set is a curated set of inputs with reviewed reference answers or scoring criteria. It gives a repeated AI task a shared standard. Label whether cases are used for development or held out for final testing; a curated dataset is not automatically an independent test.

  • Guardrails

    Guardrails are the rules and checks around an AI system that keep it within safe, allowed behavior: input filters, output checks, permission limits and human approvals. They do not make a model perfect; they limit the damage when it is wrong or manipulated.

H

  • Hallucination

    A hallucination is when an AI model states something false or made-up as if it were true, such as a fake statistic, quote or citation. It happens because models generate plausible text, not verified facts. Reduce it by giving sources, allowing 'I don't know', and checking claims before use.

  • Human-in-the-loop

    Human-in-the-loop means a person reviews, approves or corrects an AI workflow at defined points. It is one control alongside permissions, validation and monitoring. A useful review gives the person enough evidence, time and authority to change or stop the action.

I

  • Incrementality testing (lift studies)

    Incrementality testing estimates the causal effect of a campaign by comparing a treatment group with a suitable holdout. User and geographic experiments require careful design and enough data. Interpret estimated lift with uncertainty; an attribution dashboard alone does not establish what the ads caused.

  • Inference

    Inference is using a trained AI model to produce an output. A chat reply or agent task may involve several model calls. Usage, model choice and serving setup affect cost and speed; API billing can include tokens, media, tools or compute.

K

  • Knowledge cutoff

    A knowledge cutoff describes the recency of a model's built-in training knowledge, not a guarantee that it knows every earlier fact or nothing later. Current information can be supplied through search, files or tools without changing the model's weights. Check the exact model and verify time-sensitive claims.

L

  • Large language model (LLM)

    A large language model is an AI system trained on very large amounts of text to predict the next piece of text. That simple skill lets it draft, summarize, translate, classify and work through problems. Claude, GPT, Gemini and Grok are examples of LLM families. ChatGPT and similar apps are products built on top of them.

  • llms.txt

    llms.txt is a community proposal for a Markdown guide to a website's useful content. Version 2 supports a site root or a narrower path. Some documentation tools and agents use it, but publishing one does not establish a search ranking or citation benefit. Google does not require it for AI features.

M

  • Model Context Protocol (MCP)

    MCP is an open protocol for connecting AI applications to outside tools and data. A server can support multiple assistants, but transport, authentication, permissions and feature compatibility still need checking.

  • Model tiers (flagship, balanced, fast)

    Model families offer different price and capability options. Published standard API input rates for GPT-6 Astra and Luna differ by 100× before cache, long-context and processing-tier adjustments. Choose by measured task quality, latency and total cost; a tier name alone does not determine performance.

  • Multi-model aggregator

    A multi-model aggregator is a service that gives access to many AI models from different vendors through one account, interface or API. Examples include Gab AI, Perplexity and API routers such as OpenRouter. The business benefit is choice and flexibility; the cost is an extra party in your data path.

  • Multimodal AI

    Multimodal AI can take in or produce more than one kind of data, such as text, images, audio and video, in the same system. In practice, you can show a model a screenshot, chart or PDF page and ask about it, or have it reply with speech. Most frontier model families are now multimodal to some degree.

O

  • OAuth

    OAuth lets an application obtain authorized access to another service. Common account-connection flows send you to the provider for sign-in and consent, then give the app scoped tokens. OAuth is authorization, not by itself an identity protocol or a guarantee that access is narrow or safe.

  • OCR and document layout

    OCR recognizes text in an image; layout analysis identifies structure such as paragraphs, columns and tables. Both affect whether extracted content is usable. Check reading order, important fields and omissions against the source before feeding the result to an assistant.

  • Open-weight models

    Open-weight models make trained parameters available for download under a license. Examples include Mistral Small 4 and Microsoft's Phi models. Self-hosting can give you more control over data and deployment, while leaving hardware, security, maintenance and evaluation with your team.

P

  • POAS (profit on ad spend)

    POAS compares profit before advertising cost with ad spend; ROAS compares revenue with ad spend. Define which costs your profit measure includes. At a 40% contribution margin before ads, revenue ROAS of 2.5× covers those variable costs and advertising, before any excluded overhead or profit target.

  • Prompt caching

    Prompt caching reuses processing of eligible repeated input, such as instructions or reference documents. Cache hits can reduce input cost and latency, but minimum lengths, configuration, write charges and retention differ by model. Put stable content first and verify actual cache usage.

  • Prompt injection

    Prompt injection attempts to make an AI system follow attacker-controlled instructions instead of its authorized task. It can arrive directly in a message or indirectly through retrieved pages, documents, images or tool results. OWASP lists it as LLM01 in its 2025 Top 10. Tool access can increase the consequences.

R

  • Rate limits

    Rate limits cap API request or token throughput over a period. Limits vary by model, project, organization and account tier. Temporary throttling needs bounded retries; exhausted spend caps or billing problems need separate handling. Consumer-app allowances are separate.

  • Reasoning models

    Reasoning models can allocate computation to internal reasoning before and between answers or tool calls. More effort can help difficult tasks, with possible latency and token costs. The benefit is task-dependent, so compare quality, time and cost on your own examples.

  • Retrieval-augmented generation (RAG)

    Retrieval-augmented generation supplies external information to a model when it answers. Retrieval can select passages from a library larger than the context window, using keyword, vector or hybrid search. Relevance, permissions and source freshness still need checks; retrieved text does not guarantee a correct answer.

  • Rubric for scoring AI output

    A rubric defines what a reviewer should check, how each criterion is scored and what counts as acceptable. Use concrete examples and test the rubric on real outputs. Review disagreement before automating the grading; a model judge is another measurement method, not the final authority.

S

  • Shadow AI

    Shadow AI is workplace use of AI without the organization’s approval or oversight. It can expose data or create unreviewed outputs. Inventory the actual tools and tasks, then provide approved options, clear data rules and a practical way to request new capabilities.

  • Structured data (schema markup)

    Structured data is code on a web page that labels its content in a standard vocabulary, usually Schema.org, so search engines understand it: this is a product, this is its price, this is an FAQ. It can make pages eligible for rich results. Google says it must match the visible text and is not required for AI features.

  • Structured outputs

    Structured outputs constrain a supported model's response to a defined schema. They reduce formatting and parsing problems, but valid structure does not make the values correct or an action authorized. Applications must handle refusals, incomplete responses and unsupported schema features.

  • Synthetic data

    Synthetic data is artificially generated data designed to resemble useful properties of real data. It can support tests and training, but may preserve source bias or sensitive details. Validate privacy and task coverage rather than assuming that generated means anonymous or representative.

  • System prompt

    A system prompt supplies instructions that guide a model's role, behavior and response style. Products may add their own instructions alongside user preferences and tool definitions. These instructions influence responses but do not guarantee compliance or enforce access permissions.

T

  • Temperature

    Temperature is a setting that controls how random a model's output is. Low values make answers more focused and repeatable, and higher values make them more varied. It has traditionally been tuned low for extraction and higher for brainstorming. Some newer models are built to run at their default, so check the vendor's guidance.

  • The lethal trifecta (agent security)

    The lethal trifecta is Simon Willison’s name for combining private-data access, untrusted input and an outbound communication path in an AI agent. It highlights a data-leak risk. Removing a capability or enforcing strict boundaries can reduce exposure; other prompt-injection harms still need controls.

  • Tokens

    Tokens are the units models process. Text tokens may be words, fragments or punctuation, while media uses model-specific accounting. Token counts determine many context limits and API charges. Count your actual request: English word estimates do not transfer reliably across languages, media or tokenizers.

  • Tool use (function calling)

    Tool use lets a model request operations such as retrieving an order or sending a message. Your application executes client-defined functions; providers can execute hosted tools. The surrounding system controls authorization, validation and execution. A structured request is not proof the action is appropriate.

  • Turn detection (end-of-turn)

    Turn detection is how a voice agent decides that you've finished speaking and it's time to answer, or that you've started talking again and it should stop. Good turn detection makes voice AI feel natural. Bad detection interrupts people or leaves awkward silences. Some speech models now build it in, such as Deepgram's Flux and OpenAI's semantic VAD.

V

  • Vector database

    A vector database stores embeddings for similarity search. It can support semantic retrieval and RAG, but is not required for every project. Compare it with keyword search or a small context-based approach using actual token volume, permissions, retrieval quality and operating cost.

W

  • Webhooks

    A webhook is an event notification one service sends to another, often as an HTTP request. It can start an automation after a payment, form submission or other event without continual polling. Delivery can be delayed, duplicated, reordered or fail, so the receiver needs verification and recovery controls.

Z

  • Zero-click search

    A zero-click search ends without a result click, such as when the session ends or the person searches again. SparkToro and Datos estimated 58.5% of US Google searches in their 2024 study. That historical panel estimate is not a current universal rate or proof that AI answers caused the behavior.

Other directories: Needs & problems · Questions & how-tos · Services & tools · AI models · Use-case stacks