Reference
Glossary
Definitions of the core terms and concepts used throughout HelpStack. Each entry links to the guide or integration doc that covers it in depth.
Agent (role)#
A team member whose role is AGENT. Agents can handle conversations (reply, assign, change status, add notes) but have read-only access to settings. See ../guides/09-team-and-roles.md.
Agent tool#
A custom AI ability you define so the model can take actions beyond replying. A tool has one of seven types: SERVER_SIDE (HelpStack calls your HTTPS endpoint), CLIENT_SIDE (runs in the visitor's browser via the widget), EMAIL_INQUIRY (the AI emails someone outside the conversation and picks the thread back up when they reply — see its own entry below), CALLBACK_JOB (the AI hands a job to one of your systems and picks the thread back up when that system reports it finished — see its own entry below), CONTROL (built-in system tools that ship with HelpStack, whose descriptions an admin can customize), and SHOPIFY / WOOCOMMERCE (added automatically when you connect a store, on any plan). Each tool has a name, a description the LLM reads to decide when to call it, and JSON-Schema parameters. Tools can be imported from an OpenAPI spec, and every call is logged. See ../guides/08-agent-tools.md.
AI provider#
A configured LLM connection — DeepSeek, OpenAI, Anthropic, Google, xAI, Qwen, Mistral, Moonshot, MiniMax, or Ollama — with a model, temperature, and max-tokens setting. Every new organization starts with a DeepSeek provider named "DeepSeek Pro" set as its default, so AI replies work before you configure anything. One provider is the org default; the system prompt is set per channel, not on the provider. See ../guides/07-ai-providers.md.
AI reply#
An AI-drafted reply to an inbound message. Its context includes recent conversation history, prior closed conversations with the same customer, attached knowledge-base groups (RAG), the channel system prompt, the customer identifier, and any agent tools. A drafted reply waits in the response queue for agent approval unless the channel is set to auto-approve. See ../guides/03-ai-replies.md.
Assignment#
The act of attaching a conversation to a specific user, making them responsible for it. Notification scope ASSIGNED_ONLY limits alerts to conversations assigned to you. See ../guides/02-conversations.md.
Assignment rule#
A rule that routes a new conversation to a specific user automatically, based on conditions you set. Configured at Settings → Routing. A rule's assignee also receives the "Needs human" alert when the AI escalates a conversation the rule routed. See ../guides/13-routing-and-handoff.md.
Auto-reply (channel mode)#
A per-channel setting controlling how the AI drafts replies on inbound messages. Modes: off (no automatic drafting), prepare (AI drafts automatically, agent approves before sending), and auto-approve (AI drafts and sends without agent review). See ../guides/03-ai-replies.md.
Channel#
A connected source of conversations. Types: EMAIL, FACEBOOK, INSTAGRAM, WHATSAPP, WEBSITE_CHAT, LINKEDIN, and VOICE. LINKEDIN and VOICE are gated behind per-organization feature flags, so they only appear in the add-channel dialog once enabled for your organization. (TELEGRAM exists in the schema but is not yet available in the UI.) See ../guides/04-channels.md.
Conversation#
A customer thread that groups messages within a channel. A conversation has a status (OPEN → PENDING → CLOSED, with ARCHIVED available to file it away without it sitting in your closed list), can be assigned to a user, carries internal notes, and has a detected language. Both CLOSED and ARCHIVED end the conversation as far as the AI is concerned. See ../guides/02-conversations.md.
Default provider#
The single AI provider marked as the organization's default. It is used for AI replies and translation unless overridden. See ../guides/07-ai-providers.md.
Email inquiry (EMAIL_INQUIRY)#
An agent tool whose answer does not arrive in the same reply. The AI writes an email itself to someone outside the conversation, typically a carrier or a supplier, tells the customer you are checking, and picks the thread back up when the reply comes in. The conversation sits in WAITING in the meantime, and the exchange with that third party is shown on a card under the customer's thread. Off by default for every organization: it needs the Growth plan and HelpStack switching the feature on, because it sends mail from your own domain on the AI's initiative. See ../guides/08-agent-tools.md.
Async job (CALLBACK_JOB)#
An agent tool that hands a unit of WORK to one of your own systems and waits for it to report back, rather than answering inside the reply. The AI calls the tool, HelpStack POSTs the job to your endpoint, the conversation parks, and your system calls HelpStack back when the work is done or has failed.
The sibling of an email inquiry: same waiting machinery, different counterparty. An email inquiry asks a PERSON a question and waits hours, and a classifier decides whether the reply actually answered it. An async job asks a SYSTEM to do something, waits seconds to an hour, and that system states its own outcome, so nothing has to judge it.
Deadlines are set in seconds (5 to 3600). There are no automatic retries: a
repeated request could repeat the side effect, and only the receiving system
knows whether that is safe. Behind the deferredTools feature flag. See
../integrations/async-tools.md.
Escalation / handoff#
The AI handing a conversation off to a human, marking it "Needs human" and stopping its own replies until someone picks it up. Fires the ESCALATED notification event. See ../guides/13-routing-and-handoff.md.
FAQ#
A question-and-answer pair (Question ≤280 characters, Answer ≤2000 characters), defined org-wide or per WEBSITE_CHAT channel. FAQs appear as clickable chips in the widget greeting. They are distinct from the knowledge base: FAQs power the widget greeting, while the knowledge base grounds AI replies. See ../guides/06-faqs.md.
Feature flag#
A per-organization entitlement controlled by HelpStack staff, used to gate a capability that is not yet generally available (for example the LinkedIn and Voice channels). A flag you do not have simply hides the feature; there is nothing to switch on yourself. See ../guides/04-channels.md.
HMAC signature#
A cryptographic signature sent with incoming webhook payloads (for example from Facebook/Instagram/WhatsApp) that HelpStack verifies to confirm the request is authentic and unmodified. See ../integrations/webhooks.md.
Inbox#
The unified view of all conversations across every channel, where agents read, filter, assign, and reply. See ../guides/02-conversations.md.
Internal note#
A private comment added to a conversation for your team only. Notes are never sent to the customer. See ../guides/02-conversations.md.
Knowledge base#
A collection of articles used to ground AI replies through RAG. Articles are MANUAL, URL-scraped, or FILE-uploaded, and move through statuses PROCESSING → READY/ERROR. Articles are organized into groups that you attach to channels. See ../guides/05-knowledge-base.md.
Knowledge-base group#
A named set of knowledge-base articles. Attaching a group to a channel makes those articles available to the AI as RAG context for that channel's replies. See ../guides/05-knowledge-base.md.
Message#
A single message in a conversation. Direction is INBOUND (from the customer) or OUTBOUND (from you). A message may carry translatedContent and an AI-generated draft awaiting approval. See ../guides/02-conversations.md.
Notifications#
Alerts for the events NEW_CONVERSATION, NEW_MESSAGE, ASSIGNED, STATUS_CHANGED, and ESCALATED, delivered via IN_APP, PUSH, EMAIL, SLACK, TEAMS, SMS, WEBHOOK, or DISCORD. Each event has a per-channel toggle and a scope (ALL or ASSIGNED_ONLY). See ../guides/10-notifications.md.
Organization#
The tenant root. Every piece of data — channels, conversations, users, providers — is scoped to an organization. Users belong to one organization with a role. See ../guides/01-getting-started.md.
RAG (Retrieval-Augmented Generation)#
The technique of retrieving relevant knowledge-base articles and supplying them to the LLM so AI replies are grounded in your own content rather than the model's general knowledge. See ../guides/05-knowledge-base.md.
ResponseQueue#
The queue holding AI-drafted replies awaiting agent approval. When a channel is set to auto-approve, drafts skip the queue and send automatically. See ../guides/03-ai-replies.md.
Role#
A user's permission level within an organization: OWNER (full access plus billing), ADMIN (full access minus billing actions), AGENT (handle conversations, read-only settings), or VIEWER (read-only conversations and analytics). A platform-level SUPERADMIN exists for HelpStack staff. See ../guides/09-team-and-roles.md.
Scope (ALL / ASSIGNED_ONLY)#
A notification setting that controls which conversations trigger alerts. ALL notifies on every matching event in the organization; ASSIGNED_ONLY notifies only for conversations assigned to you. See ../guides/10-notifications.md.
System prompt#
The per-channel instruction that shapes the AI's tone, persona, and rules when drafting replies for that channel. Set on the channel, not on the AI provider. See ../guides/03-ai-replies.md.
Top-up#
A one-time purchase of additional tokens, separate from a recurring plan. Like all billing actions, top-ups are OWNER-only. See ../guides/12-billing.md.
Translation#
Automatic language handling: HelpStack detects the customer's language, translates inbound messages into the agent's language and outbound messages into the customer's language. Short messages (fewer than 6 words) inherit the conversation language. Performed internally via OpenAI/Anthropic and/or Google Cloud Translation. See ../guides/02-conversations.md.
VisitorId#
A unique identifier the widget persists in the browser's localStorage to recognize a returning website visitor across sessions. See ../integrations/widget-embed.md.
Voice channel#
A channel type where an AI voice agent answers spoken questions through an embeddable voice widget, rather than in text. Gated behind a per-organization feature flag. See ../guides/04-channels.md.
Verify token#
A shared secret used during platform webhook setup (Facebook, Instagram, WhatsApp) to confirm HelpStack owns the endpoint when the provider sends a verification challenge. Set via environment variables such as FACEBOOK_VERIFY_TOKEN, WHATSAPP_VERIFY_TOKEN, and INSTAGRAM_VERIFY_TOKEN. See ../integrations/webhooks.md.
Waiting (conversation status)#
WAITING means the conversation is waiting on someone outside your team, and needs nobody. It is set automatically when the AI asks a third party a question through an email inquiry tool, and it clears by itself when the answer arrives or the wait runs out. Distinct from PENDING, which means a human must act now; if a customer asks for a person while a thread is waiting, it becomes PENDING and the request for a human wins. Shown as "Waiting on Reply" in the inbox filter and not selectable from the status menu, since nothing about it is your decision. See ../guides/02-conversations.md.
Widget#
The embeddable website chat, loaded with <script src="https://helpstack.eu/widget.js?id=CHANNEL_ID" async></script>. It persists a visitorId in localStorage, communicates in real time over Socket.IO, and exposes programmatic control via window.ChatWidget.{init, open, close, toggle, identify}. Client-side agent tools are registered with window.HelpStack('registerTool', name, handler). See ../integrations/widget-embed.md.