Reference sheet · all lessons

Glossary

One name per idea. When several words exist, the course picks one and lists the others as aliases to avoid.

tawk.to

Widget: The tawk.to chat UI loaded from https://embed.tawk.to/{propertyId}/{widgetId}. Avoid: “plugin”, “chat box”.
Property: A tawk.to site/workspace with its own settings, agents, widgets and add-ons. IDs: Administration → Overview.
JavaScript API: window.Tawk_API: settings, methods and callbacks exposed by the widget in the browser. Avoid: “tawk SDK” (there is none).
Secure mode: Identity verification for the widget: a server-side HMAC-SHA256 hash of the visitor identifier with the property’s JavaScript API key. Avoid: “login token”, “JWT”.
Webhook: A signed POST from tawk.to to your URL for one of four events; HMAC-SHA1 in X-Tawk-Signature; retried for 12 hours with a stable X-Hook-Event-Id.
Direct Chat Link: https://tawk.to/chat/{propertyId}/{widgetId}, a standalone chat page that cannot be pre-filled.
AI Assist: tawk.to’s paid AI add-on; its default bot persona is “Apollo”. Answers inside the widget from its own data sources. Avoid: “Apollo” as the product name.
Data source (AI Assist): FAQs, text, documents, tables, shortcuts, Knowledge Base articles, websites; retrieval selects up to 5 chunks per inquiry.
Instructions (AI Assist): The agent’s prompt, formerly “Base Prompt”. Where scope restrictions live in Pattern B.
Knowledge Base: tawk.to’s hosted help centre (name.tawk.help); public published articles feed AI Assist and can appear as widget cards.
Shortcut: A canned response (/name) with optional files, forms, suggested messages and actions (tag, transfer). AI Assist can send one verbatim when its “AI Questions” match.
Trigger: An automatic message based on conditions (up to 20 per property); Advanced triggers can match words in the visitor’s message.
Escalation (AI Assist): Paid-plan setting that binds “asks for a human” and “no response found” to a shortcut with a Transfer Chat action.

This course’s own names

Pattern A / B / C: A: your bot is the front door, tawk.to is the human channel. B: AI Assist is the front door, your API is its tool. C: no bot.
Handoff: Moving a visitor from your bot to a human in tawk.to with identity and context attached. Avoid: “escalation” for this (reserved for AI Assist’s setting), “transfer”.
Tier: One stage of the answer pipeline: 0 human requested, 1 FAQ, 2 approved sources, 3 fallback.
FAQ tier: Deterministic matching of a visitor question to a curated entry; returns the canonical answer verbatim. Avoid: “intent matching”.
Threshold / margin: The FAQ tier’s two dials: minimum top score; minimum gap to the runner-up.
Approved sources: The one or two places answers may come from, expressed as an allow-list of hosts (optionally with a path prefix). Avoid: “whitelist”.
Grounded: A model answer with text and at least one citation into an allowed chunk. Anything else is a fallback.
Sentinel: The exact string NOT_IN_SOURCES the model returns when the documents do not answer; detected by code, not by reading prose.
Reason code: The string on every Answer that explains the decision; your analytics key.

Retrieval and Claude

Chunk: A piece of a source document stored with its URL and title; the unit of retrieval and citation.
Citation: An API-returned pointer (cited_text, document index, location) into a document block that had citations: { enabled: true }.
Domain filter: allowed_domains / blocked_domains on Claude’s web search and web fetch tools; mutually exclusive; no scheme; subdomains included.
Prior-context rule: Web fetch only fetches URLs that already appeared in the conversation (user messages, client tool results, earlier results), never URLs only in the system prompt.
Prompt caching: Reusing a stable prefix (tools → system → messages) across requests via cache_control; verify with cache_read_input_tokens.
tf-idf cosine: Lexical similarity that weights rare words more; bounded 0..1; the FAQ tier’s default.
BM25: The standard lexical ranking function (MiniSearch implements BM25+); the step up from tf-idf for larger FAQs.
Embeddings: Vectors for semantic similarity; Anthropic has no embedding model and recommends Voyage AI.
Contextual retrieval: Prepending a short document-aware context to each chunk before indexing; Anthropic measured large reductions in retrieval failure.
Graph RAG: “RAG, where the Retrieval path includes a knowledge graph.” Local search for entity-centred questions; global search for corpus-wide ones.
Triple: (subject, predicate, object) with provenance; the atom of the course’s graph prototype.
Community report: Microsoft GraphRAG’s LLM-written summary of a cluster of entities (found with hierarchical Leiden); the unit global search reads.
Multi-hop question: One that needs two or more facts joined across chunks; where graphs earn their keep.

Vercel and Android

Vercel Function: A serverless handler; this course uses the web-standard export async function POST(request: Request) form.
Raw body: The exact request bytes (await request.text()), required for signature verification.
Idempotency key: A delivery id stored so retries are ignored; for tawk.to, X-Hook-Event-Id.
Shell (Android): What wraps your bot on a phone: a Trusted Web Activity, a Capacitor app, or a native Kotlin app.
WebView chat screen: The in-app page that loads the tawk.to widget, identifies the visitor, and bridges callbacks to native code.