Hazzy answers your customers instantly from your own documents — with the source attached to every reply, so it never makes anything up. No coding. Live today.
Live interactive preview — type your own question, or try one below.
Every query, embedding, and tool call is scoped to your account at the database layer — never inferred by the model, never trusted from the client. [1]
Any action that changes a customer's account or money requires a person on your team to approve it. The assistant proposes; it never executes alone. [2]
Every retrieval and every tool call lands in an append-only audit log — who asked, what was retrieved, what came back, and when. [3]
Your embed key only answers for the domains you've allow-listed. A copied script tag can't quietly impersonate your brand elsewhere. [4]
The signs are everywhere:
Imagine if every one of those was answered instantly — correctly, with a source attached.
Upload your docs or point Hazzy at your site. It chunks, embeds, and indexes the content — then answers only from that material, with a source attached. Below a confidence floor, it hands off instead of guessing.
Create your account in minutes. No infrastructure to provision, nothing to install.
Each site gets its own isolated knowledge base — a client can run more than one.
Drop in PDFs and docs, or point Hazzy at your site and let it crawl.
Content is chunked, embedded, and indexed. Nothing goes live until you publish it.
One script tag, dropped anywhere HTML runs.
Visitors get grounded, cited answers. You get the full transcript.
Customers get the answer straight from your docs instead of waiting for someone to type the same reply again.
The widget answers around the clock — no time zone, no queue, no "we'll get back to you."
Every reply traces back to the same published source of truth — not whoever happens to be on shift.
Support and sales stop digging through five PDFs to find the one paragraph that actually matters.
Every answer traces back to a document. No citation, no answer — Hazzy retries once, then admits it doesn't know rather than guess.
Every query and embedding is scoped server-side to your account, checked at the database layer — never a prompt-level promise.
One script tag, rendered in an isolated frame so your site's styles and Hazzy's never collide. WordPress, Shopify, or raw HTML.
Once connected, the widget can look up a logged-in visitor's own order or ticket status — scoped to that one person, never anyone else's.
NewThe assistant can propose an action — replying to a ticket, say — but someone on your team confirms it before anything changes.
Every retrieval and tool call lands in an append-only audit trail — tenant, actor, input, output, timestamp.
Embeds anywhere — it's just a script tag
Connects to your backend, live today
WhatsApp and Slack are on the roadmap — not live yet. See where things stand.
Answer plan, pricing, and renewal questions from your own docs — Haznox is tenant #1, running this exact setup.
Connect WooCommerce or Shopify so the widget can look up a customer's own order — scoped to that one account.
Turn your help docs and FAQ into a widget that answers support and pre-sales questions from what you've actually published.
Client A's knowledge base and conversations must never be reachable from Client B's widget — under any prompt, any query, any edge case.
Within one client's site, one visitor's order or ticket data must never surface to another visitor through the assistant.
We'd rather ship a narrow product that works than a broad one that doesn't. Each phase only starts once the one before it is solid.
Upload, train, embed, chat — grounded and cited, on your own site.
Collections, versioning, and review workflows for larger knowledge bases.
Deeper backend connectors and account-aware answers, approval-gated throughout.
Multi-step actions on your behalf — still never without a human's sign-off.
Proactive notifications and end-to-end workflows across your tools.
| Feature | Hazzy AI | Generic AI chatbot |
|---|---|---|
| Cites its sources on every answer | Always | Rarely |
| Tenant data isolated at the database layer | Enforced | Varies |
| Says "I don't know" below a confidence floor | Always | Rarely |
| Human approval required before any write action | Enforced | Uncommon |
| Every retrieval and tool call audit-logged | Always | Rarely |
| Embeds anywhere with one script tag | Yes | Varies |
Comparing against the general category, not a specific named product — we haven't tested every alternative ourselves.
Every plan includes the full platform — upload, train, embed, chat, with a citation on every answer. Create your account, then pick a plan from your dashboard. No sales call required.
Need something larger or with specific compliance requirements? Talk to us about enterprise.
It says so. Below a confidence threshold, Hazzy hands off instead of guessing — a wrong "I don't know" costs you less than a wrong answer with your name on it.
No. Every table is scoped by tenant at the query layer, not just checked in application code — and it's verified with a deliberate adversarial test that tries to breach it.
Only the ones you turn on, and only after someone on your team approves each one individually. It can propose an action; it can't execute one unsupervised.
Inside an isolated frame on your page, loaded by a small asynchronous script. Your site's styles and scripts never touch it, and it never touches yours.
Yes — sign up and you're live today, no waitlist or sales call required.
Your content is split into chunks, turned into embeddings, and stored per-account. When someone asks a question, Hazzy retrieves the most relevant chunks — by meaning and keyword — and the model is only allowed to answer from what it retrieved, with a citation attached.
Every response is checked for a citation before it's sent. No valid citation triggers one retry with a stricter prompt, then a canned "I don't know" rather than ever shipping an uncited claim. It's a structural check, not just an instruction to the model.
Postgres for structured data and embeddings (with pgvector), object storage for raw uploaded files and transcript exports. Every table is partitioned by tenant, indexed on it, and scoped at the query layer — not just filtered after the fact.
Yes — PDF, DOCX, and plain text upload directly. Nothing reaches the live widget until you publish it, so you can review what got ingested first.
Yes — point it at one page, or let it crawl same-site links up to 20 pages and 3 levels deep. It respects robots.txt and is guarded against fetching internal/private addresses.
Brand color, greeting message, and tone today, plus up to four static conversation-starter buttons. Avatar/logo upload and widget positioning are on the near-term roadmap, not yet built.
The assistant itself will generally answer in whatever language it's asked in, since that's native to the underlying model — it isn't a separately built or tested feature yet. The widget's own buttons and labels are English-only today.
Claude is primary; if it's unreachable, the runtime automatically fails over to OpenAI or Gemini when those are configured, before ever falling back to a canned "I don't know." Provider failure degrades the answer quality, not the uptime.
Free, Starter, Growth, and Business plans, priced by message volume, knowledge base size, and seats — see current pricing above. Free never expires but sleeps after 30 days idle; paid tiers get annual billing for 2 months free, and you can change plans any time from your dashboard.