Resources · Glossary
Every term on this page appears somewhere on this site. Each entry says what the term means in our work, why it matters when you buy or run an AI system, and where on the site the idea is put to work. Search it, filter it by topic or jump by letter.
80 termsWritten by Vijay KUpdated
Each entry has three parts. The definition says what the term means when we use it, which is sometimes narrower than the general meaning. "Why it matters" is the question to ask a vendor, including us, before you sign. "Where we use it" links to the page that shows the idea working, so you can see the mechanism rather than take the word on trust.
Where a term names an outside rule, such as WhatsApp messaging windows or India’s data protection law, the entry gives the idea in a sentence and the official source is listed at the end of the page. Our summaries are not legal advice.
80 terms
Also called: application-to-person messaging on US local numbers
The US carrier registration system for business text messages sent from ordinary local numbers. A business registers its brand and each campaign, describing the purpose and how people opt in and out, before sending.
Why it matters: Unregistered business texts in the US can be filtered or blocked. Registration needs lead time in a project plan.
Also called: client-owned accounts, no lock-in
Our default: hosting, model APIs, messaging and telephony are registered to your business, keys are issued from your accounts, and we work through named roles you can revoke. A short pilot on our accounts is the agreed exception.
Why it matters: If the vendor owns the accounts, leaving the vendor means rebuilding. Ownership is the simplest protection against lock-in.
Also called: MiniAI Agent Engine
Our name for the delivery architecture behind every chatbot, agent and workflow we build: layers for knowledge, reasoning, reach and control, deployed separately for each client. It is not a product you sign up for.
Why it matters: Shared architecture means each build starts from tested parts, while your deployment, data and accounts stay yours.
Also called: agentic system
A system that works through a multi-step task, deciding which tool to use next, reading the result and continuing until the task is done or it needs a person. Our agents propose actions; code checks them and consequential ones wait for approval.
Why it matters: Agents earn their cost when the steps vary from case to case. When the steps are always the same, a plain workflow is cheaper and easier to trust.
Also called: AI assistant, conversational assistant
A system that answers people in a conversation, on a website, WhatsApp or another channel. The ones we build answer from your content, cite it and hand the conversation to a person when they are not sure.
Why it matters: A chatbot that cannot say "I don’t know" or reach a human creates support work instead of removing it.
Also called: human approval step
A point where a proposed action, with its arguments and the evidence behind it, waits for a named person to accept, edit or reject it. We put gates on refunds, cancellations, deletions and messages to many people.
Why it matters: It keeps a person accountable for consequential actions while the system still does the preparation work.
Also called: interruption handling
Letting a caller interrupt the agent mid-sentence: the agent stops speaking, listens, and responds to what was just said. It needs the audio stream and the conversation state to be handled together.
Why it matters: People interrupt on the phone all the time. An agent that talks over them sounds like an old phone menu.
Also called: keyword ranking, full-text search
A long-established formula for ranking documents by the words they share with a query, weighting rare words above common ones. It is the keyword half of our retrieval and finds exact matches such as plan names, error codes and SKUs.
Why it matters: A purely semantic search can return a near-miss for "Plan B2". Exact-match ranking keeps product codes and policy names reliable.
Also called: BAA
Under the US HIPAA rules, the contract a healthcare provider or health plan puts in place with a vendor that handles protected health information on its behalf, setting out how that information is safeguarded.
Why it matters: If a system will touch patient data for a US covered entity, the BAA question comes before the build. We review requests case by case.
Also called: channel connector
The piece that translates one channel, such as website chat, WhatsApp, a phone call or email, into one internal event format, and renders replies back within that channel’s rules. Routing, retrieval and guardrails are written once behind it.
Why it matters: Adding a channel later then means a new adapter, not a second assistant with its own knowledge and its own mistakes.
Also called: document splitting
Cutting each document into passages small enough to search and quote. We split on the document’s own headings, so a chunk is a section a person would recognise, and every chunk keeps its source, anchor and version.
Why it matters: Bad splits produce answers that quote half a rule. Good splits let a citation land on the exact section a reader can check.
Also called: source attribution
A link from a claim in an answer back to the passage it came from. In our builds the model marks each claim with a source number, and a check on the way out confirms that every cited source was actually retrieved.
Why it matters: Citations let a customer or a reviewer verify an answer in one click, and they make a wrong answer easy to trace to the document that caused it.
Also called: worked example
A worked example of how we would build a system for a type of business, with its architecture, a step trace on sample data and typical targets. It is labelled on every card and page and is never client work.
Why it matters: It shows method and depth honestly, without borrowing credibility from clients who did not agree to be named.
Also called: confidence threshold, low-confidence handoff
A rule that checks how strong the evidence behind an answer is before it is sent. When retrieval finds little, or the classifier is unsure, the system asks a clarifying question or offers a person instead of guessing.
Why it matters: It decides whether an unsure assistant guesses or admits it. The assistant on this site uses one: ask it about something this site does not cover.
Also called: context length
The amount of text a model can consider in one request: instructions, retrieved passages, conversation history and its own reply together. Anything outside the window does not exist for that answer.
Why it matters: A large window is not a substitute for retrieval. Stuffing it raises cost and can bury the passage that matters.
Also called: service level agreement
A service commitment, such as response times for issues and what is monitored, written for one client’s system and scope. We do not publish a standard uptime figure; a custom SLA is available on Enterprise builds.
Why it matters: A standard figure on a website says little about your system. A written commitment for your scope can be checked.
Also called: WhatsApp messaging window, message template
On the WhatsApp Business Platform, the period that opens when a customer messages or calls your business and resets with each new message. Inside it the business can send free-form replies; once it closes, only pre-approved, paid template messages.
Why it matters: An assistant that ignores the window fails to deliver its replies. We enforce it in the channel adapter, not in the model.
Also called: DPA
A contract that sets out how a processor may handle personal data for a client: purpose, security measures, sub-processors, breach notice and deletion at the end. We review and sign one on request, case by case.
Why it matters: Privacy laws in many markets expect one. Send yours before work starts so it shapes the build, not the other way round.
Also called: processing location
Where data is stored and processed, by country or region. Most of a system’s data sits in your own systems and in the providers it calls, so residency is chosen provider by provider where they offer regions.
Why it matters: Some sectors and contracts require data to stay in a region. Raise it at scoping, because it narrows the provider list.
Also called: retention period
How long each kind of data is kept before it is deleted. On this site, a chat conversation stored on the server expires after thirty days, and a visitor can reset or delete it from the chat window.
Why it matters: Data you no longer need is a risk with no benefit. Every build should have a written retention period per data type.
Also called: dead letters, parked jobs
Where a job goes when it has used up its retries, kept with its full payload, every attempt and the last error. A person can correct it, replay it with its original idempotency key, or discard it.
Why it matters: Without one, failed runs vanish into logs nobody reads. With one, every failure has an owner and a next step.
Also called: dedupe, record matching
Recognising that two records or two deliveries are the same thing, by a delivery id, an email address or a set of matching fields, and keeping one. Uncertain matches go to a review queue rather than being merged silently.
Why it matters: Duplicate leads and contacts erode trust in a CRM faster than missing ones.
Also called: Digital Personal Data Protection Act, 2023 (India)
India’s law on processing digital personal data. It sets duties for the data fiduciary, the business that decides why and how personal data is processed, including notice, consent for a stated purpose and deletion when the purpose ends.
Why it matters: For Indian customers it shapes intake flows, consent wording and retention. We design for it and document the choices for your privacy lead.
Also called: vector embedding
A list of numbers that a model produces for a piece of text so that passages with similar meaning sit close together. Searching embeddings finds a passage about "cancelling an order" when the visitor typed "how do I stop my purchase".
Why it matters: Embeddings catch paraphrases that keyword search misses. They also miss exact codes and names, which is why we pair them with keyword search.
Also called: eval set, test set, golden set
A fixed collection of test cases written from your real questions and workflows, each with the expected behaviour: the right source, a correct refusal or a handoff. Every change runs against it offline, in-process, and before release against a deployed preview.
Why it matters: Without a written set, "it seems better" is the only test. With one, a release either passes or it does not.
Also called: model tuning
Further training a model on your examples so it adopts a format, tone or labelling habit. It changes how a model behaves, not reliably what it knows, and it cannot cite where a fact came from.
Why it matters: Fine-tuning is rarely the first step. Retrieval handles facts that change; tuning is a measured later step for behaviour prompting cannot fix.
Also called: fixed price, fixed scope
An engagement where scope, price and timeline are agreed in writing before work starts, with a working demo before final payment. Changes are quoted as changes, not absorbed silently.
Why it matters: It puts the estimation risk on the builder and gives you a clear definition of done.
Also called: grounded answer
Keeping an answer inside the evidence the system found. A grounded reply says only what the retrieved passages or a live lookup support, and says "I don’t have that" when they do not.
Why it matters: Grounding is the difference between an assistant that admits a gap and one that fills it with something plausible.
Also called: AI safety controls
The layered controls around a model: untrusted input handling, tool allowlists, approval gates, output checks and handoffs. No single guardrail makes a system safe; together they reduce what can go wrong and limit what a mistake can do.
Why it matters: Ask for the list of layers, not a promise. Each one should be something you can see working.
Also called: confabulation, unsupported answer
A fluent answer that is not supported by any source, such as an invented policy, price or feature. No model is free of it; systems reduce it by grounding answers, checking citations and handing off when evidence is weak.
Why it matters: Any vendor who says their AI never makes things up is overselling. Ask instead how a wrong answer is caught and what the user sees.
Also called: project handover
The end of a build: the code, configuration, evaluation set and runbook are delivered, your team is walked through operating it, and our access is removed unless an operate engagement continues.
Why it matters: A system you cannot run without the builder is a dependency, not an asset. Ask what you receive on the last day.
Also called: HIPAA-aware build
Our wording for builds designed with HIPAA’s safeguards in mind: minimum necessary data, health data kept in systems you control or with providers under your own agreements, and testing on de-identified data. It is not a certification.
Why it matters: No vendor can make your organisation compliant. Your compliance advisor makes the final call, and our job is to make that review easy.
Also called: escalation, talk to a human
Passing a conversation from the assistant to a person, with the transcript, the detected intent and the sources the system looked at, so the customer never repeats themselves. When a visitor hands a chat on this site to us, the stored enquiry gets a reference starting ML-.
Why it matters: A handoff that loses context frustrates customers more than no assistant at all. Ask what your team receives.
Also called: HITL, human oversight
A design where people review, approve or take over at defined points instead of the system acting alone. It covers approval gates, handoffs and scheduled transcript review.
Why it matters: It is the practical form of oversight that governance frameworks ask for, and the reason a mistake reaches a person before a customer.
Also called: hybrid retrieval, reciprocal rank fusion (RRF)
Running keyword search and vector search on the same question and merging the two ranked lists. We merge them by reciprocal rank fusion: each passage scores by its position in each list, so no tuning between very different score scales is needed.
Why it matters: Most wrong answers start with the wrong passage. Two retrievers, merged predictably, make it less likely that the right one is never seen.
Also called: dedupe key
A unique key sent with a write, derived from the business event (for example the order and the refund request), so that repeating the request has the same effect as sending it once. The receiver recognises the key and returns the first result.
Why it matters: It is what stops a retry from charging a card twice or creating a second lead. The contact form on this site uses one.
Also called: our own system
A system we built and run on miniailabs.in itself, such as the assistant on this site and the lead pipeline behind the contact form. You can try it, and it carries the label "In-house build".
Why it matters: It is evidence you can check yourself, which is worth more than a screenshot.
Also called: router, intent classification, state machine
Classifying each message into a closed list of intents before any expensive step runs. Greetings, requests for a person and injection attempts follow fixed paths with no model call; the rest go to the specialist for that intent.
Why it matters: Routing first keeps costs down and makes behaviour predictable: each intent is a named state with one tested path, and explicit rules decide the transitions.
Also called: supersede on change, content hash
Giving every document a stable id and a fingerprint of its content, so that a changed document replaces its old passages in the same run that adds the new ones. Removed documents leave the index too.
Why it matters: Without it, an assistant can quote last year’s refund policy next to this year’s. Ask any vendor how an old version leaves the index.
Also called: LLM, foundation model
A model trained on large amounts of text to predict and generate language. In our systems it is one component: it reads, classifies and drafts, while ordinary code decides what it may see and what it may do.
Why it matters: Treating the model as one replaceable part, not the whole product, is what lets you change providers without rebuilding.
Also called: minimum necessary access, tool allowlist
Giving each person, tool and integration only the access its job needs. In our builds each step may propose only the tools on its allowlist, and every connected system gets its own scoped credential, issued from an account in your name.
Why it matters: When one credential is exposed, least privilege decides whether you lose one capability or everything.
Also called: read-only tool
Fetching a fact that changes by the minute, such as an order status, a free slot or a balance, from the system that owns it at the moment of the question, instead of copying it into the knowledge base.
Why it matters: A stale copy of live data is worse than none. A lookup answers from the record your team also sees.
Also called: model-graded evaluation
Using a language model to grade another model’s answers against a rubric, for checks that exact matching cannot do, such as whether a reply stayed on its sources. The judge itself is checked against human grades before its scores are trusted.
Why it matters: An uncalibrated judge can agree with itself all day. Ask how often its grades were compared with a person’s.
Also called: model per step, provider fallback
Choosing which model serves each step, such as a small fast model for classification and a stronger one for drafting, behind one internal interface. A fallback route takes over when a provider is slow or down.
Why it matters: One model for everything usually costs more and fails harder. Routing chosen by evaluation keeps quality, speed and cost in balance.
Also called: AI RMF
A voluntary framework from the US National Institute of Standards and Technology for managing AI risks, organised around four functions: govern, map, measure and manage. It has no certification scheme: an organisation aligns its own process with it.
Why it matters: It gives buyers and vendors a shared vocabulary for risk reviews without inventing a new checklist.
Also called: monitoring
Being able to see what a running system is doing from what it records: run logs, error counts, retries, handoffs and failures waiting in the dead-letter queue. We describe what we watch instead of publishing an uptime figure.
Why it matters: A system that fails quietly is the expensive kind. Ask what you will see, and who is alerted, when a step breaks.
Also called: self-hostable model
A model whose trained weights are published so you can run it on your own servers or cloud account, such as Llama or Mistral models. Licences differ by model and must be read before commercial use.
Why it matters: It lets sensitive steps run without data leaving your environment, at the cost of running and updating the model yourself.
Also called: messaging consent
A person’s recorded agreement to receive messages from your business on a channel, naming your business, collected before you message them. We store when and how it was given, and honour opt-outs immediately.
Why it matters: Consent rules differ by channel and country, and a missing record is the first thing a platform or regulator asks for.
Also called: orchestration layer
The ordinary code around the model that routes each request, calls the right specialist step, runs tools, handles retries and decides when to stop. In our Agent Engine it is the Reason layer.
Why it matters: Most of an AI system’s reliability lives here, in code you can test, not in the model’s good behaviour.
Also called: reply checks
Code that checks every reply before it is sent: citations must point at retrieved passages, prices must match published ones, links must be on an allowlist and claims the business cannot make are removed.
Why it matters: It catches the specific mistakes that would embarrass you, such as an invented price or a certification you do not hold.
Also called: filter before ranking, access-controlled search
Applying who may see what inside the search itself, so a passage a user is not allowed to read is never retrieved and never reaches the model. It matters most for internal assistants over wikis and shared drives.
Why it matters: Filtering after the answer is written is too late: the model may already have paraphrased the restricted text.
Also called: personal data masking
Removing personal and secret data, such as card numbers, government ID numbers, passwords, one-time codes and API keys, from messages before they are stored, logged or sent to a model.
Why it matters: Data that is never stored cannot leak. Redaction early in the pipeline shrinks everything downstream that needs protecting.
Also called: indirect prompt injection
Text written to make a model ignore its instructions, typed by a user or hidden in a document, web page or tool result the system reads. We treat all such text as data, scan for it at ingestion and retrieval, and keep consequential actions behind approval.
Why it matters: There is no complete fix today; design reduces the risk and limits the damage. Be wary of anyone who claims otherwise.
Also called: reference design
A worked design for a common system, such as a support chatbot or an inbound voice agent, showing its components, data flow, failure modes and cost drivers. A real build starts from one and adapts it to your systems.
Why it matters: It lets you see the whole system, and its weak points, before a line of code is written.
Also called: release gate
A rule in the release pipeline that blocks a deploy when any gated metric on the evaluation set falls below its threshold, or any critical case fails. Thresholds are set from the first good baseline.
Why it matters: Model and prompt changes can fix one thing and break another. The gate catches the break before customers do.
Also called: RAG, grounded generation
A pattern where the system first searches your own content for passages relevant to a question, then asks the model to answer using only those passages. The model writes the sentence; your documents supply the facts.
Why it matters: It is how an assistant can answer from a policy that changed this morning and show where the answer came from, without retraining any model.
Also called: exponential backoff, jitter
Repeating a failed call after a wait that grows with each attempt, plus a small random delay so queued jobs do not all hit a recovering service at once. We retry only transient failures such as timeouts and rate limits.
Why it matters: Retrying a permanent error, such as a rejected field, just repeats the failure. Knowing which errors are worth retrying is half of reliability.
Also called: exception queue
A list where items the system is unsure about wait for a person: an extracted field below its confidence threshold, a conflicting CRM record, a proposed action. Each item shows what the system saw and what it suggests.
Why it matters: It turns "the AI might be wrong" into a bounded, visible amount of human work instead of silent errors.
Also called: execution log, audit log
A structured record of each conversation or automation run: the trigger, each step with a summary of its input and output, timing, attempts, status, and the model and prompt version used. Personal data is redacted first. Admin changes go to a separate audit log.
Why it matters: When something goes wrong, the run log is how your team finds out what happened without asking an engineer.
Also called: illustrative data
Invented records, names and messages used in demos, traces and screenshots on this site. Every such panel carries a visible "Sample data" label, and people appear as roles, never as realistic company names.
Why it matters: A demo that looks like real customer data invites the wrong conclusion. The label keeps the illustration honest.
Also called: OAuth scopes, restricted API key
An access token or API key limited to named permissions, such as read contacts and write notes, often granted through OAuth so a person approves exactly what is shared. It can be revoked without touching anything else.
Why it matters: Ask which scopes each integration requests and why. Broad admin keys are a convenience you pay for later.
Also called: discovery call, strategy call
The free first call where we map one workflow, its systems and its risks, and agree what a first build should and should not do. A written proposal with a fixed scope and price follows.
Why it matters: Most project risk is removed here, by deciding what is out of scope before anyone builds.
Also called: on-premise, private deployment
Running every layer of a system, models included, on infrastructure you control, such as your cloud account or your own servers. A hybrid variant runs only the sensitive steps on a self-hosted open-weight model.
Why it matters: It keeps data inside your environment at the price of operating the stack. It is usually an Enterprise-scope decision.
Also called: responsibility split
A written split of who handles which control: what we build and configure, what each provider runs, and what stays with you, such as user access, content accuracy and regulatory sign-off.
Why it matters: Gaps appear where everyone assumes someone else is responsible. A table with names closes them.
Also called: ASR, transcription; text-to-speech (TTS) is the reverse
Turning a caller’s speech into text as they talk, so the voice agent can understand the request. Streaming recognition returns partial text while the caller is still speaking.
Why it matters: Recognition errors on names, numbers and accents are where voice agents most often go wrong. Test with your real callers’ speech.
Also called: JSON schema output
Asking a model to return a typed object that matches a schema, such as an intent label with a confidence and the fields it extracted, instead of free prose. Code then validates the object before anything acts on it.
Why it matters: Prose is hard to check; a typed object either passes validation or it does not, which makes a pipeline testable.
Also called: third-party processor
A provider that processes personal data on behalf of the business responsible for it, such as a hosting platform, a model API or an email service. Each proposal we write names every provider in the data path.
Why it matters: You cannot assess risk in providers you do not know about. The list is the starting point of any security review.
Also called: source of truth
The one system whose copy of a fact counts when two disagree: the CRM for a contact, the store for an order, the payment provider for a charge. Automations read from it and write back to it rather than keeping their own copy.
Why it matters: Naming the system of record for each object up front prevents two tools from overwriting each other.
Also called: instructions, system message
The standing instructions a system gives the model before the conversation: its role, the rules it follows and how to treat the sources it is handed. In our builds it is versioned like code and lives on the server, never in the browser.
Why it matters: Whoever controls the system prompt controls the assistant’s behaviour. It should be reviewed, versioned and kept out of reach of visitors.
Also called: SIP, phone numbers, call routing
The phone side of a voice agent: numbers, inbound and outbound calls, transfers to a person and recordings, provided through a platform such as Twilio and connected to the agent by a media stream.
Why it matters: Telephony carries its own costs, consent rules and number registration, separate from the AI. Budget and plan for it on its own line.
Also called: tokens
The unit a language model reads and writes, roughly a short word or part of a word. Providers bill and limit usage in tokens, counting the prompt that goes in and the reply that comes out.
Why it matters: Token use drives the running cost of an assistant. A prompt that carries a whole manual on every turn costs more than one that retrieves the right section.
Also called: function calling, tool use
Letting a model propose a call to a defined function, such as "look up order" with its arguments, which the surrounding code validates and runs. The model never runs anything itself.
Why it matters: Tools are where an AI system touches your business. Each one needs a schema, a scope and a rule for when a person must approve it.
Also called: outbox pattern
Writing a record and the notifications it should trigger in the same database transaction, then sending the notifications from that stored list with retries. The lead pipeline on this site stores a lead and its emails together this way.
Why it matters: It means a lead is never saved without its alert, and an alert is never sent for a lead that failed to save.
Also called: response latency, latency budget
The pause between the caller finishing and the agent starting to reply, made up of recognition, retrieval or tools, the model and speech synthesis. We plan a budget for each stage before building.
Why it matters: On a call, long pauses read as a dropped line. Each stage has to earn its share of the wait.
Also called: design target
How we present an outcome number: a range for a named metric with the basis it rests on, such as an evaluation set or your own ticket history. It is a goal the build is designed and measured against, never a result claimed for a client.
Why it matters: An outcome number without a stated basis is marketing. A target with a basis can be measured after launch.
Also called: vector database, vector store
The store that holds embeddings and returns the nearest ones to a query. We use Postgres with pgvector through Supabase, or Cloudflare Vectorize, chosen by where the rest of the system runs.
Why it matters: It is one more place your data lives. It should sit in an account in your name and be cleaned when a source document is removed.
Also called: HTTP callback
A message one system sends to another’s web address when something happens, such as a payment succeeding or a WhatsApp message arriving. We verify the sender’s signature, validate the payload and drop duplicates before acting.
Why it matters: Senders retry webhooks they think failed, so the same event can arrive twice. A receiver must be built for that.
Also called: WhatsApp Cloud API
Meta’s programmatic access to WhatsApp for businesses, as opposed to the WhatsApp Business phone app. A system sends and receives messages through the API on a number registered to your business account.
Why it matters: The number, the business account and the message history should be in your name, not your vendor’s.
Also called: process automation
Connecting business tools so a trigger, such as a new lead or a paid invoice, runs a fixed sequence of steps without someone copying data between screens. An AI step may classify or draft inside it; the sequence itself is set.
Why it matters: For repeatable work it is cheaper and easier to trust than an agent, and it is where most first projects pay back.
Also called: sync back
Recording the result of an AI step in the system your team already works in, such as a qualified lead’s notes in the CRM or a ticket tag in the helpdesk, through that system’s API.
Why it matters: If the result stays inside the AI tool, your team has a second place to check. Write-back keeps one place.
FAQ
We do. Each definition is written by MiniAI Labs in the sense we use the term in our builds, not copied from a vendor or a dictionary. Where a term names an outside rule, the official source is listed at the end of the page.
A definition tells you what a word means; the linked page shows the idea working in a real design, with its limits. If a term has no live page yet, its name is shown without a link until that page is published.
It starts the same way, but retrieval for an assistant runs two kinds of search at once, filters by what the person may see, and hands only the best passages to the model, which must cite them. A help-centre search box shows the reader a list; retrieval decides what the model is allowed to say.
No. A typical target is the range a build is designed and measured against, with the basis printed next to it. It is not a result claimed for any client, and the Work page shows concept builds, not client projects.
Because we hold no certifications and compliance is decided by your organisation and its advisors, not by a vendor. "Aligns with" and "HIPAA-aware" describe how we design; they are not a certificate, and we will not use words that suggest one.
Yes, send it to hello@miniailabs.in with the context you met it in. We add terms that come up in our own builds and buyer conversations, and we date the page when we change it.
Yes. The glossary is part of what the assistant answers from, so it can explain a term and link the definition. If it is not sure, it says so and offers to pass your question to a person.
Next step
Book a free 30-minute strategy call. Bring one process you want to automate and we will talk through scope, integrations and a sensible first step.
Free 30-minute call · no commitment
