Architecture · how it works
How GrootDesk thinks — explained simply.
No jargon required. Below we walk through exactly what happens when a customer asks a question — first in plain English, then in as much technical depth as you want.
In plain English
Think of it as your best support agent — who never sleeps.
A great human agent does five things: understands the question, checks the order, looks up your policy, writes a clear reply, and calls a manager only when they truly need to. GrootDesk does exactly that — automatically, in seconds, for every customer at once.
It listens
Reads the customer's message and works out what they actually want.
It checks
Verifies who they are and looks up their real order from your shop.
It looks it up
Finds the answer in your own policies, FAQs and product info.
It replies
Writes a clear, correct answer — and asks permission before doing anything.
It knows its limits
If a human is needed, it hands over neatly — with the whole story attached.
The journey of one question
This is the actual decision flow every message travels through. Follow the arrows.
💬 Customer sends a message
Step 1 · Understand
Work out what they want
One quick check figures out the intent, whether an instant action fits, and which model to use. (Node: intent classifier)
🔀 What does this need?
Needs order data
→ look up the live order first
Just a policy / FAQ
→ search the knowledge base
Needs a human
→ skip ahead to hand-off
Step 2 · Check (only if needed)
Look up the live order & product data
Pulls real, up-to-date data straight from your shop — after verifying the customer. (Node: data fetcher)
Step 3 · Look it up
Search your knowledge base
Finds the most relevant policy, FAQ or document so the answer is grounded in your words. (Node: RAG retriever)
Step 4 · Reply
Write the answer — live
Composes a clear reply grounded in what it found, streamed word-by-word so the customer sees it forming. (Node: response generator)
🛟 Can it safely finish on its own?
Yes
→ send the answer (and, with consent, complete the action)
No
→ create a ticket & hand off to a human, context attached
(Node: escalation detector)
✅ Done — and every step is logged to your dashboard
The moving parts
Three pieces work together: your shop, the GrootDesk brain, and the outside services it uses — all backed by one secure, EU-hosted data store. Your shop and the brain talk over a secure live connection (WebSocket), and your shop proves who each visitor is with a signed token (JWT).
Your shop
what your customer sees
- Chat widget
- Plugin / script snippet
- Signed session token (JWT)
- Store & order API
GrootDesk brain
where the thinking happens
- Live connection (WebSocket)
- Unified classifier
- LangGraph pipeline
- Knowledge search + cache
- Analytics dashboard
Outside services
the tools it plugs into
- LLM providers (any)
- Helpdesk (Zendesk, Freshdesk)
- Automations
- Email / notifications
One secure memory — hosted in the EU
The pipeline, node by node
Under the hood, GrootDesk runs a LangGraph state machine — a set of small, single-purpose steps with clear rules between them. Because each step is separate and traceable, nothing is a black box.
-
01
Intent Classifier
Decides what the message needs and routes it the right way.
-
02
Data Fetcher
Pulls live orders, products & customer data from your store.
-
03
RAG Retriever
Searches your knowledge base for the most relevant facts.
-
04
Response Generator
Writes the grounded answer and streams it back live.
-
05
Escalation Detector
Decides whether to finish, or hand off to a human.
Long chats are quietly condensed by a summarizer so context stays sharp without ballooning cost. Simple, deterministic requests (tracking, invoice, lookup) skip the pipeline entirely and are answered instantly — no model call, no cost.
Instant actions
These are handled directly — no AI needed — so they're instant, free and always reliable.
How we keep it honest
Every function is traced and every answer is graded — so quality is measured, not hoped for.
- Full tracing — every step is recorded and replayable.
- Hallucination check — confirms the answer is backed by your content.
- Language match — always replies in the customer's language.
- Conciseness score — keeps replies short and clear.
The dashboard
One screen to watch everything.
You're never in the dark. A live dashboard shows every conversation, what it cost, what customers asked about, what got escalated — and lets you tune the agent without touching code.
Explore the dashboard →14
views, one place
Live
conversations & logs
€/chat
cost per conversation
No-code
tune agents & prompts
Built with
Give your team its time back.
Join the early-access programme — founder pricing locked for 12 months, live on your store in under an hour.