Breezee AI

Appendix A — Glossary

Terms used in this manual, alphabetised.

TermDefinition
ActionA tool the agent can call during a conversation — display a form, render a card, embed a Cal.com widget. Auto-instantiated by skill configuration; there's no standalone Actions UI in the current product.
ActivityA first-class event on a prospect record: chat session, meeting booking, contact form submission, profile change, integration sync. Visible on the Prospects screen's History tab.
AgentThe AI chatbot. Configured with a profile (persona, profession, response length, goal), a greeting, a set of skills, and routing rules. One team can have multiple agents.
Agent typeA template that determines an agent's default skill set and prompts. The only generally-available agent type today is sAIlsbot; others are placeholders.
AttributeA structured property extracted from product-shaped content (e.g. waterproof_rating, plan_tier). Used by the Product Finder skill. Not relevant for services-led tenants.
Book Meeting skillThe skill that surfaces an inline Cal.com booking widget in the chat. Configured per agent with a Cal.com event URL.
Cal.comThe booking provider the platform integrates with for Book Meeting. The widget loads as an iframe pointed at a Cal.com event URL.
ChannelA deployment target type. Currently Chat widget is the only available channel.
ChunkA self-contained passage of ingested content with its own vector embedding, used for retrieval. The chunker is AI-driven and tries to preserve semantic coherence.
Consent promptThe widget's GDPR-aware "Remember me / No thanks" dialog. Fires after personal data is shared (booking, contact form) or after four messages.
Contact Request skillThe skill that displays an inline contact form and submits captured details to the prospect record.
ContentIngested knowledge — scraped URLs or uploaded documents. Each piece of content moves through parse → classify → chunk → attribute extract → embed before being usable by the agent.
Conversation QualityA 1–5 score from the summariser estimating how well the agent handled a session. Averaged across sessions on the Analytics screen.
Custom memory typeA memory type you defined yourself (rather than one of the platform's seeded built-ins). Lives in the Custom tab of the Memory screen.
DeploymentA configured instance of an agent on a channel. Has its own theme, content copy, allowed domains, and embed snippet. One agent can have multiple deployments.
Deployment keyA unique UUID that identifies a deployment. Used in the widget loader URL.
Discourage ruleA routing rule that steers the agent away from a target skill when criteria match. Stateless; all matching discourage rules fire on every turn. Can optionally provide a Suggest instead alternative.
Embed snippetThe single-line <script> tag you paste into a website to install the widget. Bound to a specific deployment via the deployment key.
FirecrawlThe third-party web-scraping service used to map and scrape URLs during the Import URLs flow.
GreetingThe first message the agent sends, plus the suggested-question chips and chat-input hint that appear with it. Configurable per agent and overridable per deployment.
Informational skillThe skill that answers questions about the company, services, policies, etc. using the knowledge base. Includes an FAQ priority fast-path for high-similarity FAQ chunks.
InsightsThe right-hand panel in the playground that fills in after a session ends — summary, sentiment, outcome, resolution, quality, content analysis. Mirrors what the summariser computes for every session.
IntegrationA connected third-party service (HubSpot today). Authenticated via OAuth 2.0; managed from the team's Integrations screen.
Knowledge baseAll the content ingested into a team — the corpus the agent retrieves from during conversations.
Lead ScoreA 0–100 score the platform computes from a prospect's memories, profile matches, and activities. Higher = warmer.
Manage dataThe visitor-facing panel for reviewing and erasing personal data, reachable from the widget's menu icon. The platform-side equivalent of a GDPR data-subject access request.
MemoryA specific value the agent captured about a prospect (e.g. Company Name = Breezee).
Memory typeThe schema for a kind of memory the agent can capture (e.g. Company Name, Current Challenge). Configured per team in the Memory screen. Either built-in or custom.
MQLMarketing Qualified Lead. A built-in prospect profile name; the criteria that define it are tenant-configurable.
OrganisationThe top-level tenant. Everything in the platform belongs to an organisation.
PlaygroundThe in-dashboard test environment for trying conversations against your agent. Same chat pipeline as the production widget.
ProfileSee Prospect profile.
Prompt materialisationThe server-side process of filling prompt templates with team / agent / deployment values, run automatically when configuration changes.
ProspectA visitor who has interacted with the chat widget. Each prospect carries the captured memories, the conversation history, and any contact information they shared.
Prospect profileA named pattern of memory criteria that classifies a prospect — MQL, Hot Lead, Tire Kicker, etc. Profiles are consumed by routing rules.
Qualify Prospect skillThe skill that actively captures the structured memories that drive qualification. Configured with a per-field mode (Always ask, Ask when natural, Infer silently).
Reciprocal Rank Fusion (RRF)The scoring technique the retrieval engine uses to combine vector similarity, title match, and content text match into a single ranked result list.
Remember meThe visitor's affirmative answer to the consent prompt. Once accepted, in-session memories persist to the prospect record on session end.
Resolution RateThe percentage of sessions the platform classified as Resolved — i.e. ones where the visitor reached a meaningful outcome.
Retrieval policyOne of default, product_specific, or informational. Determines how the search engine weights vector / title / content scoring and whether the FAQ fast-path applies.
Routing engineThe deterministic, non-LLM component that evaluates routing rules against the current session's slot state every turn and injects directives into the responder prompt.
Routing ruleA user-configured Suggest or Discourage rule that steers the agent toward or away from a target skill when prospect-profile criteria match.
SessionA single chat conversation. Created when a visitor opens the widget or starts a playground session; closed by End session or a long idle period.
Session actions menuThe three-dot menu in the playground that holds Manage data and End session.
SkillA modular AI capability you attach to an agent. Out of the box: greeting, informational, product_finder, advisory, book_meeting, contact_request, order_form, qualify_prospect, plus a few defensive ones.
Slot extractorThe LLM call that pulls structured values (memories) out of the visitor's message on every turn. Runs in parallel with intent and tool planning.
Suggest ruleA routing rule that steers the agent toward a target skill when criteria match. Has a per-session lifecycle (pending → fired → resolved). First match wins among suggest rules.
SummariserThe background LLM call that runs after each chat turn and at session end. Updates the running session summary and (at end) produces the full Insights output.
TeamA workspace inside an organisation. Holds agents, content, prospects, and deployments. The URL paths call this concept projects; the UI calls it team.
Tool plannerThe LLM call that decides per-turn whether to search the knowledge base and which retrieval policy to use.
Unstructured.ioThe third-party document-parsing service used to extract text from uploaded PDFs, Word docs, and similar non-Markdown formats.
WidgetThe embeddable chat interface that visitors interact with on your website. Renders as an iframe inside a launcher button.