Every field of helppuff.json. Generated from the schema helppuff validates against, so it is always exact; helppuff schema prints the same as JSON Schema. For what these files are and how they relate to the dashboard, see Configuration.

Secrets are never values here: a field marked { env } takes the name of an environment variable, e.g. { "env": "OPENAI_API_KEY" }, whose value lives in .env and on the Worker.

Field	Type	Default	Description
$schema	string		The JSON Schema for editors and agents. Written by helppuff.
format	integer		The helppuff.json format version. Absent means 1; helppuff upgrade updates it when a release changes the format. · ≥ 1
site (required)	string		Site id: lowercase letters, digits and dashes. Appears in the embed snippet.
name (required)	string		The business name, as visitors see it. · 1–60 chars
website	string		The website the assistant is for, and learns from. · URL
origins (required)	string[]		Every origin the widget may be embedded on. The preview page is added automatically. · ≥ 1 items
model	object		Who writes the answers. Leave it out for Workers AI. Changed only here (or with helppuff model set), then helppuff deploy.
model.provider (required)	"workers-ai" | "openai-compatible" | "openai" | "gemini" | "anthropic" | "custom"		"workers-ai" Workers AI, on your Cloudflare account. The default: no key, on the Free plan. · "openai-compatible" Any OpenAI-compatible /chat/completions API with tools. · "openai" OpenAI. · "gemini" Google Gemini (its OpenAI-compatible API). · "anthropic" Anthropic Claude (the Messages API). · "custom" Your own model: a TypeScript file whose default export is defineModel({ id, chat }). See the wiki's Custom model page.
model.model	string		A Workers AI model id. Default: GLM-4.7 Flash. · 1–200 chars
model.budget	object		Spend guards. Other providers bill you directly; cap them with security.limits.messagesPerSitePerDay.
model.budget.dailyNeurons	integer		Workers AI only: neurons per UTC day before answers stop (default 9,000; the free allocation is 10,000). Raise it on Workers Paid. · 0–10000000
model.budget.maxInputTokens	integer		The most tokens sent to the model per answer: prompt, passages and history, trimmed to fit. · 1000–100000
model.reasoning	"low" | "medium" | "high"	"medium"	How long the model thinks before answering: low, medium or high. Deeper is better on multi-step questions, slower to start, and the thinking is billed. Models that only switch thinking on or off treat every level as on.
model.fallbackModel	string		Tried once when the main model fails for any reason other than the budget. · 1–200 chars
model.locale	string		BCP 47, e.g. en-AU: spelling and date style of answers. · ≤ 35 chars
model.timezone	string		IANA, e.g. Australia/Melbourne: "are you open now?". · ≤ 64 chars
model.maxAnswerSentences	integer	4	The longest answer, in sentences. Short answers cost less and read better in a chat. · 1–20
model.maxOutputTokens	integer	600	A hard cap on tokens per answer. · 64–4096
model.historyMessages	integer	6	Earlier messages sent with each question. The main cost lever after retrieval. · 0–24
model.richMessages	boolean	true	Let the model offer next-step chips (option buttons) after an answer.
model.tools	object	{}	What the assistant can do besides answering.
model.tools.callback	boolean	true	Offer to have the team call or email back — the default way to a person.
model.tools.businessHours	boolean	true	Answer "are you open now?" from the business hours and time zone.
model.tools.captureLead	boolean		Retired. Accepted from older configs and ignored.
model.tools.handoff	boolean		Retired. Accepted from older configs and ignored.
model.tools.booking	boolean		Retired. Accepted from older configs and ignored.
model.business	object	{}	Details the owner confirmed. Usually left empty: the details learned from the site, and edited in the dashboard, are used.
model.business.name	string		The business name. · ≤ 200 chars
model.business.phone	string		The main phone number. · ≤ 100 chars
model.business.email	string		The contact email. · ≤ 200 chars
model.business.address	string		The street address. · ≤ 400 chars
model.business.hours	string[]	[]	Opening hours, one line per day or range, e.g. "Mon-Fri 7am-5pm". · ≤ 14 items
model.business.serviceAreas	string[]	[]	Suburbs or regions served. · ≤ 100 items
model.gateway	string		With provider: "workers-ai". An AI Gateway id: caching, logs and rate limits in front of every Workers AI call. · 1–64 chars
model.preset	"deepinfra" | "openrouter" | "deepseek" | "groq" | "together" | "mistral" | "fireworks" | "vercel-ai-gateway" | "cloudflare-ai-gateway"		With provider: "openai-compatible". Fills baseUrl and the key's variable name: deepinfra, openrouter, deepseek, groq, together, mistral, fireworks, vercel-ai-gateway, cloudflare-ai-gateway.
model.baseUrl	string		With provider: "openai-compatible". The API base, before /chat/completions. Needed without a preset. · URL
model.apiKey	{ env }		With provider: "openai-compatible". The key, by environment variable name. Default: the preset's (e.g. DEEPINFRA_API_KEY).
model.headers	map of string | { env }		With provider: "openai-compatible". Extra headers; a value may be a secret ({ env }).
model.nativeTools	boolean		With provider: "openai-compatible". The model calls tools natively (default true). Off: tool calls it writes in its text are still read.
model.accountId	string		With provider: "openai-compatible". cloudflare-ai-gateway: the account (default: the one deployed to).
model.gatewayId	string		With provider: "openai-compatible". cloudflare-ai-gateway: the gateway (default: default). · 1–64 chars
model.gatewayToken	{ env }		With provider: "openai-compatible". cloudflare-ai-gateway: an authenticated gateway's token, sent as cf-aig-authorization.
model.apiKey	{ env }	{"env":"OPENAI_API_KEY"}	With provider: "openai". Your OpenAI API key, by environment variable name.
model.apiKey	{ env }	{"env":"GEMINI_API_KEY"}	With provider: "gemini". Your Gemini API key, by environment variable name.
model.apiKey	{ env }	{"env":"ANTHROPIC_API_KEY"}	With provider: "anthropic". Your Anthropic API key, by environment variable name.
model.module (required)	string		With provider: "custom". The file, relative to helppuff.json, e.g. ./llm.ts. Bundled into the Worker at deploy.
model.secrets	string[]		With provider: "custom". The environment variables your file reads (env.NAME): uploaded as Worker secrets at deploy. · ≤ 20 items
prompt	string	"prompt.md"	Path to the system prompt, relative to helppuff.json.
assistant	object	{}	How the assistant behaves: goal, tone, answer length. HelpPuff writes these around prompt.md on every answer, so prompt.md holds only what is specific to the business.
assistant.goal	"callbacks" | "answers" | "bookings"	"callbacks"	What the assistant is for: callbacks (help, then get the team in touch), answers, or bookings (help, then send them to bookingUrl: a booking, sign-up or quote page).
assistant.tone	"friendly" | "professional" | "casual"	"friendly"	How it sounds.
assistant.length	"short" | "detailed"	"short"	short: a few sentences; detailed: complete answers with short lists.
assistant.prices	"share" | "quote"	"share"	share: give prices exactly as the site and documents state them; quote: never give a price or estimate, offer a quote from the team instead.
assistant.bookingUrl	string		The page the bookings goal sends visitors to: booking, sign-up or a quote form. · ≤ 2000 chars, URL
autoSummary	boolean	true	Automatically summarise and label finished conversations. Off keeps transcripts, after-chat tools and webhooks running.
live	object	{}	Live chat: visitors can talk to a person on your team, who answers from the dashboard or Telegram. Off by default. helppuff live on turns it on.
live.enabled	boolean	false	Let visitors talk to a person on the team. When nobody is available they get the callback form instead.
live.waitSeconds	integer	120	How long a visitor waits for someone to take the chat before they are offered the callback form (they can keep waiting). · 15–3600
live.closeAfterMinutes	integer	60	A conversation with no message for this long is closed. A visitor who writes again is answered by the assistant. · 5–1440
live.showAgentName	boolean	true	Show visitors the first name of the person answering ("Sam joined"); off: "Someone from the team".
live.aiWhileWaiting	boolean	false	Let the assistant keep answering until someone takes the chat.
bookings	object	{}	Appointments: mode none, link (offer a booking page) or picker (a date picker in the chat; HelpPuff books, the assistant never does). Left out, a booking link found on the website is offered. With the picker, sync picks the calendar (keys in Settings → Integrations, helppuff integrations). helppuff bookings changes them.
files	object	{}	Photos and files on R2: visitors sending them in the chat (chat: off, live while a person has the chat, or always) and the team attaching them to jobs (jobs). Both off by default; helppuff files turns them on. Needs R2 on for the Cloudflare account (deploy makes the bucket).

Only for a backend that runs the whole conversation itself: retell, http (your own API), echo, or openai / gemini with their own file stores. Otherwise leave it out and use model and knowledge.retrieval.

HelpPuff's own assistant, set up by model and knowledge.retrieval. Implied when backend is left out; never written by hand.

Field	Type	Default	Description

Workers AI with HelpPuff's own knowledge base. The default; runs on the Workers Free plan.

Field	Type	Default	Description
backend.model	string	"@cf/zai-org/glm-4.7-flash"	The Workers AI model that writes answers. GLM-4.7 Flash is the best all-round on the Free plan; the wiki's AI models page compares the others for speed, quality and answers a day. · 1–200 chars
backend.reasoning	"low" | "medium" | "high"	"medium"	How long the model thinks before answering: low, medium or high. Deeper is better on multi-step questions, slower to start, and the thinking is billed. Models that only switch thinking on or off treat every level as on.
backend.fallbackModel	string		Tried once when the main model fails for any reason other than the budget. · 1–200 chars
backend.gateway	string		An AI Gateway id: caching, logs and rate limits in front of every model call. · 1–64 chars
backend.locale	string		BCP 47, e.g. en-AU: spelling and date style of answers. · ≤ 35 chars
backend.timezone	string		IANA, e.g. Australia/Melbourne: "are you open now?". · ≤ 64 chars
backend.maxAnswerSentences	integer	4	The longest answer, in sentences. Short answers cost less and read better in a chat. · 1–20
backend.maxOutputTokens	integer	600	A hard cap on tokens per answer. · 64–4096
backend.historyMessages	integer	6	Earlier messages sent with each question. The main cost lever after retrieval. · 0–24
backend.richMessages	boolean	true	Let the model offer next-step chips (option buttons) after an answer.
backend.retrieval	object	{}	How answers find passages in the knowledge base.
backend.tools	object	{}	What the assistant can do besides answering.
backend.tools.callback	boolean	true	Offer to have the team call or email back — the default way to a person.
backend.tools.businessHours	boolean	true	Answer "are you open now?" from the business hours and time zone.
backend.tools.captureLead	boolean		Retired. Accepted from older configs and ignored.
backend.tools.handoff	boolean		Retired. Accepted from older configs and ignored.
backend.tools.booking	boolean		Retired. Accepted from older configs and ignored.
backend.handoff	map of any		Retired (callbacks replaced handoff). Accepted from older configs and ignored.
backend.business	object	{}	Details the owner confirmed. Usually left empty: the details learned from the site, and edited in the dashboard, are used.
backend.business.name	string		The business name. · ≤ 200 chars
backend.business.phone	string		The main phone number. · ≤ 100 chars
backend.business.email	string		The contact email. · ≤ 200 chars
backend.business.address	string		The street address. · ≤ 400 chars
backend.business.hours	string[]	[]	Opening hours, one line per day or range, e.g. "Mon-Fri 7am-5pm". · ≤ 14 items
backend.business.serviceAreas	string[]	[]	Suburbs or regions served. · ≤ 100 items
backend.budget	object	{}	The daily Workers AI spend guard. Past it, visitors get your contact details and a callback form instead of answers.
backend.budget.dailyNeurons	integer	9000	Neurons per UTC day before answers stop; the free allocation is 10,000. · 0–10000000
backend.budget.maxInputTokens	integer	6000	The most tokens sent to the model per answer — prompt, passages and history — trimmed to fit. · 1000–100000

How answers find passages in the knowledge base.

Field	Type	Default	Description
backend.retrieval.embeddingModel	string	"@cf/baai/bge-m3"	Turns pages and questions into vectors. Changing it re-learns the site (and uploaded files) on the next deploy. · 1–200 chars
backend.retrieval.rerankerModel	string | null	"@cf/baai/bge-reranker-base"	Re-reads the passages search found and keeps only those that answer the question. null turns it off: about half a second faster, but more likely to answer from the wrong page.
backend.retrieval.topKVector	integer	20	Passages taken from meaning (vector) search before re-scoring. · 1–50
backend.retrieval.topKKeyword	integer	20	Passages taken from keyword (full-text) search before re-scoring. · 1–50
backend.retrieval.finalK	integer	4	Passages the model reads for each answer. · 1–10
backend.retrieval.minScore	number	0.2	Passages scored below this are dropped. With none left, the assistant says it is not sure and offers a callback instead of guessing. · 0–1
backend.retrieval.queryRewrite	"heuristic" | "llm" | "off"	"heuristic"	heuristic: follow-ups borrow the previous question. llm: one extra small call rewrites it.
backend.retrieval.intentModel	string | null	null	e.g. @cf/cloudflare/clef-flash: classify each question to steer retrieval. Off by default.

Cloudflare AI Search: retrieval and generation managed by Cloudflare.

Field	Type	Default	Description
backend.instance	string		Instance name on your account. Created by helppuff deploy if it does not exist.
backend.endpoint	string		Use an existing public endpoint instead of a binding, e.g. https://search.example.com · URL
backend.model	string		Workers AI model id, or an AI Gateway alias. Empty uses the instance's own model.
backend.maxResults	integer		Passages AI Search retrieves per question. · 1–50

OpenAI (Responses API), or any compatible endpoint with baseUrl.

Field	Type	Default	Description
backend.model	string	"gpt-5-mini"	The OpenAI model.
backend.apiKey	{ env }	{"env":"OPENAI_API_KEY"}	Your OpenAI API key, by environment variable name.
backend.vectorStoreId	string		Filled in by helppuff knowledge sync.
backend.retrieval	"helppuff"		helppuff: answer from HelpPuff's own knowledge base (crawled by the Worker) instead of a vector store.
backend.promptId	string		A stored prompt in the OpenAI dashboard; overrides prompt.md.
backend.baseUrl	string		Another OpenAI-compatible Responses endpoint, e.g. Azure OpenAI. · URL

Google Gemini, with File Search.

Field	Type	Default	Description
backend.model	string	"gemini-3-flash"	The Gemini model.
backend.apiKey	{ env }	{"env":"GEMINI_API_KEY"}	Your Gemini API key, by environment variable name.
backend.fileSearchStore	string		Filled in by helppuff knowledge sync.
backend.retrieval	"helppuff"		helppuff: answer from HelpPuff's own knowledge base instead of File Search.

Anthropic Claude, grounded in AI Search or HelpPuff's own knowledge base.

Field	Type	Default	Description
backend.model	string	"claude-opus-5"	The Claude model.
backend.apiKey	{ env }	{"env":"ANTHROPIC_API_KEY"}	Your Anthropic API key, by environment variable name.
backend.effort	"low" | "medium" | "high" | "xhigh" | "max"		How hard Claude thinks before answering. Lower is faster and cheaper.
backend.knowledge	boolean		Ground answers in a Cloudflare AI Search instance. On by default when there is knowledge.
backend.retrieval	"helppuff"		helppuff: ground answers in HelpPuff's own knowledge base instead of AI Search.
backend.instance	string		Instance name on your account. Created by helppuff deploy if it does not exist.
backend.endpoint	string		Use an existing public endpoint instead of a binding, e.g. https://search.example.com · URL

Your own API.

Field	Type	Default	Description
backend.url (required)	string		Your API's base URL. · URL
backend.mode	"helppuff" | "openai"	"helppuff"	helppuff: your API speaks the HelpPuff backend protocol. openai: any /chat/completions endpoint.
backend.model	string		openai mode: the model name your endpoint expects.
backend.token	{ env }		Sent as Authorization: Bearer ….
backend.signingSecret	{ env }		Signs every request with HMAC-SHA256 so your API can verify it.
backend.stream	boolean	true	Stream replies as they are written.

A Retell chat agent.

Field	Type	Default	Description
backend.agentId (required)	string		The Retell chat agent id.
backend.apiKey	{ env }	{"env":"RETELL_API_KEY"}	Your Retell API key, by environment variable name.
backend.retrieval	"helppuff"		helppuff: crawl the site into HelpPuff's knowledge base; the agent searches it via a custom function (see helppuff status).

Echoes what it is sent, with a demo of every widget feature. For development; needs no key.

Field	Type	Default	Description

What the assistant learns from: the website, and your own files.

Field	Type	Default	Description
knowledge.website	boolean | object	true	Learn from the website: true (the defaults), false, or which pages and how often.
knowledge.website.maxPages	integer		Default: 300 for workers-ai (crawled by the Worker), 50 for the others (crawled here). · 1–1000
knowledge.website.include	string[]		Only crawl URLs matching one of these: a substring, or a glob like **/services/**.
knowledge.website.exclude	string[]		Never crawl URLs matching one of these. Defaults leave out privacy, terms, tags, carts and accounts.
knowledge.website.renderJs	"auto" | "always" | "never"		workers-ai: render pages drawn by JavaScript with Browser Rendering (auto = only when needed).
knowledge.website.schedule	"off" | "daily" | "weekly" | "monthly"		workers-ai: re-crawl the selected pages on this schedule.
knowledge.files	string[]	[]	Files and folders to index, relative to helppuff.json. PDF, Markdown, text, HTML, DOCX.
knowledge.retrieval	object		What answers come from. Left out: HelpPuff's own knowledge base. none for none. Changed only here (or with helppuff rag set), then helppuff deploy.
knowledge.retrieval.type (required)	"helppuff" | "none" | "ai-search" | "openai-vector-store" | "http" | "custom"		"helppuff" HelpPuff's own knowledge base: your site and files, learned by the Worker (Vectorize + D1). · "none" No knowledge base: the prompt and the business details only. · "ai-search" Cloudflare AI Search, searched for passages; your model writes the answer. · "openai-vector-store" An OpenAI vector store you fill in OpenAI, searched for passages. · "http" Your own search over HTTP: POST url { query, question, siteId, limit } → { passages: [{ title, content, url? }] }. · "custom" Your own knowledge base: a TypeScript file whose default export is defineRetriever({ id, search }). See the wiki's Custom knowledge base page.
knowledge.retrieval.embeddingModel	string	"@cf/baai/bge-m3"	With type: "helppuff". Turns pages and questions into vectors. Changing it re-learns the site (and uploaded files) on the next deploy. · 1–200 chars
knowledge.retrieval.rerankerModel	string | null	"@cf/baai/bge-reranker-base"	With type: "helppuff". Re-reads the passages search found and keeps only those that answer the question. null turns it off: about half a second faster, but more likely to answer from the wrong page.
knowledge.retrieval.topKVector	integer	20	With type: "helppuff". Passages taken from meaning (vector) search before re-scoring. · 1–50
knowledge.retrieval.topKKeyword	integer	20	With type: "helppuff". Passages taken from keyword (full-text) search before re-scoring. · 1–50
knowledge.retrieval.finalK	integer	4	With type: "helppuff". Passages the model reads for each answer. · 1–10
knowledge.retrieval.minScore	number	0.2	With type: "helppuff". Passages scored below this are dropped. With none left, the assistant says it is not sure and offers a callback instead of guessing. · 0–1
knowledge.retrieval.queryRewrite	"heuristic" | "llm" | "off"	"heuristic"	With type: "helppuff". heuristic: follow-ups borrow the previous question. llm: one extra small call rewrites it.
knowledge.retrieval.intentModel	string | null	null	With type: "helppuff". e.g. @cf/cloudflare/clef-flash: classify each question to steer retrieval. Off by default.
knowledge.retrieval.instance	string		With type: "ai-search". Instance name on your account. Created by helppuff deploy if it does not exist.
knowledge.retrieval.endpoint	string		With type: "ai-search". Use an existing public endpoint instead of a binding, e.g. https://search.example.com · URL
knowledge.retrieval.maxResults	integer		With type: "ai-search". Passages per question. · 1–50
knowledge.retrieval.vectorStoreId (required)	string		With type: "openai-vector-store". The vector store id, e.g. vs_….
knowledge.retrieval.apiKey	{ env }	{"env":"OPENAI_API_KEY"}	With type: "openai-vector-store". An OpenAI API key that can read it.
knowledge.retrieval.url (required)	string		With type: "http". Your endpoint. · URL
knowledge.retrieval.token	{ env }		With type: "http". Sent as Authorization: Bearer ….
knowledge.retrieval.module (required)	string		With type: "custom". The file, relative to helppuff.json, e.g. ./rag.ts.
knowledge.retrieval.secrets	string[]		With type: "custom". The environment variables your file reads: uploaded as Worker secrets at deploy. · ≤ 20 items

Rate limits, daily cap, Turnstile. The defaults are safe for a public site.

Field	Type	Default	Description
security.captcha	object		Check new chats (and dashboard sign-ins) with Cloudflare Turnstile (invisible for most visitors). Strongly recommended: without it, a script can start chats.
security.captcha.provider (required)	"turnstile"		Cloudflare Turnstile.
security.captcha.siteKey (required)	string		The Turnstile site key (public). · 1–200 chars
security.captcha.secret (required)	{ env }		The Turnstile secret key, by environment variable name.
security.limits	object	{}	Per-visitor and per-site limits. The daily cap is the cost backstop.
security.signIn	object	{}	Dashboard sign-in limits.
security.signIn.attemptsPerIp	integer	10	Sign-in attempts (and one-time link checks) one IP may make per window. · 1–1000
security.signIn.attemptsPerAccount	integer	5	Failed sign-ins one email may have per window, from anywhere. Counts failures only. · 1–1000
security.signIn.windowMinutes	integer	15	The window both sign-in limits count over, in minutes. · 1–1440
security.signIn.captcha	boolean	true	Ask for Turnstile on the sign-in form when security.captcha is set. Add the dashboard's hostname to the Turnstile widget.
security.allowIps	string[]	[]	IP addresses or CIDR ranges exempt from the per-visitor limits (your office, a monitor). The per-chat and daily caps still apply. · ≤ 500 items
security.blockIps	string[]	[]	IP addresses or CIDR ranges refused by the chat. The dashboard is not affected. · ≤ 500 items
security.sessionTtlHours	number	24	How long a chat can be continued (the widget keeps it across pages and reloads). · 0.25–720

Per-visitor and per-site limits. The daily cap is the cost backstop.

Field	Type	Default	Description
security.limits.messagesPerIpPerMinute	integer	10	Messages one visitor (IP) may send a minute. · 1–600
security.limits.messagesPerIpPerDay	integer	100	Messages one visitor (IP) may send a day (UTC), so one visitor cannot use up the daily cap. · 1–100000
security.limits.sessionsPerIpPerHour	integer	5	New chats one visitor (IP) may start an hour. · 1–1000
security.limits.sessionsPerIpPerDay	integer	20	New chats one visitor (IP) may start a day (UTC). · 1–10000
security.limits.messagesPerSession	integer	60	Messages in one chat before the visitor must start another. · 1–1000
security.limits.messagesPerSitePerDay	integer	500	The cost backstop. Always set this. · 1–1000000
security.limits.maxMessageLength	integer	1000	The longest message a visitor may send, in characters. · 1–4000
security.limits.maxLeadFieldLength	integer	200	The longest answer to one form field, in characters (a message box gets maxLeadMessageLength). · 20–2000
security.limits.maxLeadMessageLength	integer	2000	The longest answer to a message box in a form, in characters. · 20–4000
security.limits.feedbackPerIpPerMinute	integer	30	Thumbs up or down one visitor (IP) may give a minute. · 1–600
security.limits.pollsPerIpPerMinute	integer	120	Checks for new messages one visitor (IP) may make a minute (backends that reply later, like Retell). · 1–600
security.limits.endsPerIpPerMinute	integer	10	Chats one visitor (IP) may close a minute. · 1–600
security.limits.retellLookupsPerMinute	integer	120	Knowledge-base lookups a Retell agent may make a minute, for the whole site. · 1–6000
security.limits.apiRequestsPerKeyPerMinute	integer	120	Requests a new API key may make a minute (each key can have its own). Chat requests also count against the daily cap. · 1–6000
security.limits.apiKeysPerSite	integer	50	Active API keys a site may have. · 1–500
security.limits.handoversPerIpPerDay	integer	3	Times one visitor (IP) may ask for a person a day (live chat). · 1–1000
security.limits.waitingPerSite	integer	20	Live chats that may wait for a person at once; past it, visitors get the callback form. · 1–1000
security.limits.liveSocketsPerIp	integer	3	Live chat connections one visitor (IP) may hold open at once. · 1–100
security.limits.slotLookupsPerIpPerMinute	integer	30	Times one visitor (IP) may load free times in the booking calendar a minute. · 1–600
security.limits.bookingsPerIpPerHour	integer	5	Appointments one visitor (IP) may book, move or cancel an hour, from the chat. · 1–100
security.limits.uploadsPerIpPerHour	integer	20	Photos and files one visitor (IP) may send an hour, when visitors can send files (files.chat). · 1–1000

Where leads go besides the dashboard.

Field	Type	Default	Description
leads.webhook	string | { env }		POST each lead to this URL (or to the URL in this environment variable). For every event, add webhooks in the dashboard instead.

The dashboard at /admin: conversations, leads, knowledge, analytics and settings, stored in D1 on your account.

Field	Type	Default	Description
dashboard.enabled	boolean	true	Serve the dashboard at /admin.
dashboard.adminEmail	string		The owner's email, for backends without a setup link. workers-ai: set on the setup page instead. · email
dashboard.summaryModel	string		Workers AI model used for conversation summaries.

Written by helppuff deploy. Safe to commit; holds no secrets.

Field	Type	Default	Description
cloudflare.accountId	string		The Cloudflare account deployed to.
cloudflare.workerName	string		The Worker's name. Default: knowtific-helppuff-.
cloudflare.url	string		Where the Worker answers. · URL
cloudflare.kvNamespaceId	string		The KV namespace (live config, new-conversation counters).
cloudflare.d1DatabaseId	string		The D1 database (conversations, leads, knowledge).
cloudflare.vectorizeIndex	string		The Vectorize index (knowledge vectors).
cloudflare.filesBucket	string		The R2 bucket (photos and files), when the account has R2.

Brand, launcher, home screen, lead form, flows — see helppuff schema.

Names, colour and theme.

Field	Type	Default	Description
widget.brand.name	string	"Chat"	The business name in the widget. · 1–60 chars
widget.brand.agentName	string	"Assistant"	The assistant's name, in its header and messages. · 1–60 chars
widget.brand.avatar	string		The assistant's picture: a square image URL. · 1–2048 chars
widget.brand.accent	string	"#5B5BF7"	The main colour (hex). Text on it is made readable automatically.
widget.brand.theme	"light" | "dark" | "auto"	"auto"	auto follows the visitor's system setting.
widget.brand.tokens	map of string		Fine-grained styling: CSS custom properties without the --hp- prefix, e.g. { "radius-panel": "12px", "font": "Inter, sans-serif" }.

The button that opens the chat.

Field	Type	Default	Description
widget.launcher.position	"bottom-right" | "bottom-left"	"bottom-right"	Which corner the button sits in.
widget.launcher.offset	object		Move the button away from the corner, e.g. above a cookie banner.
widget.launcher.offset.x (required)	number		Pixels from the side. · 0–200
widget.launcher.offset.y (required)	number		Pixels from the bottom. · 0–200
widget.launcher.label	string		Text beside the orb, or inside it when shape is pill. For example "Chat with us". · ≤ 40 chars
widget.launcher.icon	"chat" | "phone" | "mail" | "calendar" | "quote" | "pin" | "clock" | "wrench" | "heart" | "info" | "book" | "arrow-right" | "arrow-left" | "close" | "send" | "menu" | "sound" | "sound-off" | "check" | "external" | "person"	"chat"	Which of the built-in icons the launcher shows.
widget.launcher.shape	"orb" | "pill"	"orb"	orb is the signature circle. pill widens it to sit the label inside the button, which reads as a clearer invitation on a busy page.
widget.launcher.hideOnPaths	string[]		Pages where the widget does not appear, e.g. /checkout/**. · ≤ 50 items

The first screen visitors see.

Field	Type	Default	Description
widget.home.title	string	"Hi there"	The heading of the first screen. · ≤ 120 chars
widget.home.subtitle	string	"Ask anything, or pick a shortcut."	The line under it. · ≤ 240 chars
widget.home.shortcuts	object[]		Buttons on the first screen: suggested questions, a call button, a form… · ≤ 8 items
widget.home.shortcuts[].id (required)	string		Unique among the shortcuts. · 1–64 chars
widget.home.shortcuts[].label (required)	string		What the shortcut says. · 1–80 chars
widget.home.shortcuts[].description	string		A line under the label. · ≤ 160 chars
widget.home.shortcuts[].icon	"chat" | "phone" | "mail" | "calendar" | "quote" | "pin" | "clock" | "wrench" | "heart" | "info" | "book" | "arrow-right" | "arrow-left" | "close" | "send" | "menu" | "sound" | "sound-off" | "check" | "external" | "person"		A built-in icon.
widget.home.shortcuts[].action (required)	object		What happens when the button is pressed. kind picks it.
widget.home.shortcuts[].action.kind (required)	"reply" | "url" | "tel" | "email" | "flow" | "form" | "booking"		"reply" Send a message as the visitor. · "url" Open a page. · "tel" Call a number. · "email" Write an email. · "flow" Start one of widget.flows. · "form" Open one of widget.forms. · "booking" Open the booking calendar in the chat (Settings → Bookings: date picker in the chat).
widget.home.shortcuts[].action.id (required)	string		An id, unique within its list. · 1–64 chars
widget.home.shortcuts[].action.label (required)	string		What the button says. · 1–120 chars
widget.home.shortcuts[].action.value (required)	string		With kind: "reply". The message sent. · 1–500 chars
widget.home.shortcuts[].action.url (required)	string		With kind: "url". The page to open. · 1–2048 chars
widget.home.shortcuts[].action.newTab	boolean		With kind: "url". Open in a new tab.
widget.home.shortcuts[].action.phone (required)	string		With kind: "tel". The number to call. · 1–40 chars
widget.home.shortcuts[].action.email (required)	string		With kind: "email". The address to write to. · 3–200 chars
widget.home.shortcuts[].action.flowId (required)	string		With kind: "flow". The flow's id. · 1–64 chars
widget.home.shortcuts[].action.formId (required)	string		With kind: "form". The form's id. · 1–64 chars
widget.home.shortcuts[].paths	string[]		Show it only on these pages. · ≤ 20 items
widget.home.links	object		A list of useful pages on the first screen.
widget.home.links.title (required)	string		The list's heading. · ≤ 120 chars
widget.home.links.items (required)	object[]		The links. · 1–10 items
widget.home.links.items[].label (required)	string		The link text. · 1–160 chars
widget.home.links.items[].url (required)	string		Where it goes (https). · 1–2048 chars
widget.home.links.items[].description	string		A line under the link. · ≤ 300 chars

The short form before the chat.

Field	Type	Default	Description
widget.leadForm.enabled	boolean	true	Ask for details before the chat starts.
widget.leadForm.title	string		The form's heading. · ≤ 120 chars
widget.leadForm.fields	object[]	[{"name":"name","label":"Name","type":"text","required":true,"autocomplete":"name"},{"name":"phone","label":"Phone","type":"tel","required":true,"autocomplete":"tel"}]	The questions, in order. Add your own; each answer is kept on the lead and shown to the assistant. · 1–12 items
widget.leadForm.fields[].name (required)	string		The key the answer is stored under. name, email, phone and message mean something to the assistant (a message opens the chat). · 1–64 chars
widget.leadForm.fields[].label (required)	string		What the visitor sees. · 1–160 chars
widget.leadForm.fields[].type (required)	"text" | "email" | "tel" | "textarea" | "select"		The kind of input. select needs options.
widget.leadForm.fields[].required	boolean		Must be filled in.
widget.leadForm.fields[].placeholder	string		Hint text inside the empty field. · ≤ 160 chars
widget.leadForm.fields[].options	string[]		The choices of a select. · ≤ 50 items
widget.leadForm.fields[].pattern	string		A regular expression the answer must match. · ≤ 200 chars
widget.leadForm.fields[].autocomplete	string		The HTML autocomplete hint, e.g. email, tel. · ≤ 64 chars
widget.leadForm.submitLabel	string		The button's text. · ≤ 60 chars
widget.leadForm.privacy	object		A privacy note under the form.
widget.leadForm.privacy.text (required)	string		The notice, e.g. "We only use this to reply to you." · 1–300 chars
widget.leadForm.privacy.url (required)	string		Your privacy policy. · 1–2048 chars
widget.leadForm.askFirstMessage	boolean		Add a box for the visitor's first message to the form. A field named message does the same and can be labelled.

The conversation screen.

Field	Type	Default	Description
widget.chat.placeholder	string		Hint text in the message box. · ≤ 120 chars
widget.chat.initialMessages	string[]		The assistant's greeting, before the visitor writes. · ≤ 3 items
widget.chat.shortcuts	object[]		Buttons above the message box during the chat. · ≤ 8 items
widget.chat.shortcuts[].id (required)	string		Unique among the shortcuts. · 1–64 chars
widget.chat.shortcuts[].label (required)	string		What the shortcut says. · 1–80 chars
widget.chat.shortcuts[].description	string		A line under the label. · ≤ 160 chars
widget.chat.shortcuts[].icon	"chat" | "phone" | "mail" | "calendar" | "quote" | "pin" | "clock" | "wrench" | "heart" | "info" | "book" | "arrow-right" | "arrow-left" | "close" | "send" | "menu" | "sound" | "sound-off" | "check" | "external" | "person"		A built-in icon.
widget.chat.shortcuts[].action (required)	object		What happens when the button is pressed. kind picks it.
widget.chat.shortcuts[].action.kind (required)	"reply" | "url" | "tel" | "email" | "flow" | "form" | "booking"		"reply" Send a message as the visitor. · "url" Open a page. · "tel" Call a number. · "email" Write an email. · "flow" Start one of widget.flows. · "form" Open one of widget.forms. · "booking" Open the booking calendar in the chat (Settings → Bookings: date picker in the chat).
widget.chat.shortcuts[].action.id (required)	string		An id, unique within its list. · 1–64 chars
widget.chat.shortcuts[].action.label (required)	string		What the button says. · 1–120 chars
widget.chat.shortcuts[].action.value (required)	string		With kind: "reply". The message sent. · 1–500 chars
widget.chat.shortcuts[].action.url (required)	string		With kind: "url". The page to open. · 1–2048 chars
widget.chat.shortcuts[].action.newTab	boolean		With kind: "url". Open in a new tab.
widget.chat.shortcuts[].action.phone (required)	string		With kind: "tel". The number to call. · 1–40 chars
widget.chat.shortcuts[].action.email (required)	string		With kind: "email". The address to write to. · 3–200 chars
widget.chat.shortcuts[].action.flowId (required)	string		With kind: "flow". The flow's id. · 1–64 chars
widget.chat.shortcuts[].action.formId (required)	string		With kind: "form". The form's id. · 1–64 chars
widget.chat.shortcuts[].paths	string[]		Show it only on these pages. · ≤ 20 items
widget.chat.fallbackContact	object		How to reach a person when the assistant is unavailable (an outage, or the daily budget is spent).
widget.chat.fallbackContact.phone	string		Shown when the assistant cannot answer. · ≤ 40 chars
widget.chat.fallbackContact.email	string		Shown when the assistant cannot answer. · ≤ 200 chars

A message that pops up beside the button to invite a chat.

Field	Type	Default	Description
widget.teaser.text (required)	string		The message, e.g. "Need a quote? Ask me." · 1–200 chars
widget.teaser.delayMs	integer		Time on the page. Omit to rely on afterScroll alone. · 2000–120000
widget.teaser.afterScroll	integer		Percentage of the page scrolled, 1–100. Whichever trigger fires first shows the teaser; a visitor who reads rather than waits still sees it. · 1–100
widget.teaser.paths	string[]		Show it only on these pages. · ≤ 50 items
widget.teaser.oncePerSession	boolean	true	Show it once per visit, not on every page.

Guided questions asked in the widget (no AI), sent as one message at the end. Type: object[].

Field	Type	Default	Description
widget.flows[].id (required)	string		What a flow action or shortcut starts it by. · 1–64 chars
widget.flows[].steps (required)	object[]		Questions asked one at a time, in the widget, before anything is sent. · 1–10 items
widget.flows[].steps[].field (required)	string		The name the answer is kept under, for the template. · 1–64 chars
widget.flows[].steps[].ask (required)	string		The question, shown as the assistant's message. · 1–400 chars
widget.flows[].steps[].input (required)	"text" | "choice" | "phone" | "email"		How the visitor answers.
widget.flows[].steps[].choices	string[]		The buttons of a choice step. · ≤ 12 items
widget.flows[].steps[].required	boolean		The step cannot be skipped.
widget.flows[].submit (required)	object		What happens with the answers: message sends them as one message, job saves them as a job.
widget.flows[].submit.as (required)	"message" | "job"		"message" Sent as one visitor message. · "job" Saved as a job (the site's Jobs pipeline), with the answers as its fields; the visitor gets its number.
widget.flows[].submit.template (required)	string		The message, with {{field}} for each answer, e.g. "Quote for {{service}} in {{suburb}}". · 1–1000 chars

Inline forms by id, opened by a form action or shortcut. Submitting one sends its answers to the assistant. Type: map of object.

Field	Type	Default	Description
widget.forms.<id>.title	string		The form's heading. · ≤ 120 chars
widget.forms.<id>.fields (required)	object[]		The questions. · 1–12 items
widget.forms.<id>.fields[].name (required)	string		The key the answer is stored under. name, email, phone and message mean something to the assistant (a message opens the chat). · 1–64 chars
widget.forms.<id>.fields[].label (required)	string		What the visitor sees. · 1–160 chars
widget.forms.<id>.fields[].type (required)	"text" | "email" | "tel" | "textarea" | "select"		The kind of input. select needs options.
widget.forms.<id>.fields[].required	boolean		Must be filled in.
widget.forms.<id>.fields[].placeholder	string		Hint text inside the empty field. · ≤ 160 chars
widget.forms.<id>.fields[].options	string[]		The choices of a select. · ≤ 50 items
widget.forms.<id>.fields[].pattern	string		A regular expression the answer must match. · ≤ 200 chars
widget.forms.<id>.fields[].autocomplete	string		The HTML autocomplete hint, e.g. email, tel. · ≤ 64 chars
widget.forms.<id>.submitLabel	string		The button's text. · ≤ 60 chars

A soft sound when a reply arrives.

Field	Type	Default	Description
widget.sound.enabled (required)	boolean		Play it.

Set by security.captcha; you do not need to set it here.

Field	Type	Default	Description
widget.captcha.provider (required)	"turnstile"		Cloudflare Turnstile.
widget.captcha.siteKey (required)	string		The Turnstile site key. · 1–200 chars

Ignored: the widget always shows "Powered by HelpPuff". Accepted so older configs still load. Type: boolean | object.

UI string overrides. Keys are validated by the widget, not here. Type: map of string.