A support assistant connected to the team’s knowledge.
We built an AI support assistant connected to Chatwoot and an organization’s support information. The project combines knowledge retrieval, a conversation workflow and a route for human handoff.
Private note · Refund-policy question detected — escalated to a human with the full thread attached.
Retrievalover three years of past support conversations.
Chatwootanswering inside the live inbox.
Handoffto a person, with the conversation context.
The same questions, answered a thousand times — by a human, every time.
A Lithuanian conference organiser fields the same support questions endlessly — deadlines, payments, abstract status, registration details — across hundreds of events. The answers already existed in three years of past conversations. Every question still landed on a person, who had to read, recall, and retype what the team had answered before. Sending that history to a public AI tool was not an acceptable data flow: it is real customer correspondence, and hosting had to be agreed.
- ✕Every question answered by a human, from scratch
- ✕Years of answers locked in past conversations
- ✕Slow replies, especially outside work hours
- ✕Customer correspondence could not be pasted into a public chatbot
- ✓Routine questions answered from retrieved history
- ✓That history is the source material, not a fine-tuned model
- ✓A first response in the existing inbox, including outside office hours
- ✓Hosting and model choices documented for the project
Their documented answers, retrieved when a customer writes in.
Our founder built this in-house at a Lithuanian conference organiser — not as a Zyndix client delivery. The assistant retrieves from the organisation’s own conversation history and is wired into Chatwoot, the inbox the team already uses. When a customer writes in, it looks up relevant past answers and drafts a reply. Anything it is unsure about is handed to a person with the thread attached. Hosting used an open model on infrastructure the organisation controls; that was a project choice, not a universal privacy guarantee. The same kind of scope is described in our AI agent development and private AI and knowledge systems services.
Looks up how the team actually replied, then drafts from those passages.
The real thing — answering a real question.
A real customer asks how to submit an abstract — and the assistant answers inside the live Chatwoot widget, using retrieved passages from the team’s support history.
Actual conversation from the live assistant · powered by Chatwoot
From stored conversations to a reply in the inbox — in five stages.
Retrieve from real history
Three years of support conversations are indexed as source material. The assistant retrieves relevant passages when a question arrives — it is not a model fine-tuned on that history.
Answer from retrieved passages
The model is instructed to use those passages. It can still be wrong or incomplete; uncertain cases go to a person rather than being treated as solved.
Wire into Chatwoot
Connected to Chatwoot, so it works inside the inbox the team already runs.
Handle routine questions
Customer questions that match the retrieved material get a first response in natural language, including outside office hours.
Know when to step back
Anything uncertain is handed to a human — with the conversation context, not a cold transfer.
Hosting chosen for this project
Built on an open model and infrastructure the organisation controls. That was the agreed data flow. Other projects may use a provider API; nothing here is a claim that customer data never leaves a network.
Grounded in real history
Retrieval over three years of actual conversations means answers can follow how the team really responds. It is source lookup, not training a new model.
Lives in Chatwoot
No new tool to learn — the assistant works inside the inbox the team already uses every day.
Knows when to step back
Agreed rules decide what it answers and what it escalates, with full context handed to a human. That reduces unsupported replies; it does not prevent every mistake.
Always on in the inbox
A first response can go out when the widget is live, including outside office hours, in the customer’s language.
A knowledge-backed conversation in the tool the team already runs — not a generic chatbot with no source and no handoff.
3 yrsof conversations used as retrieval sources.
RAGlookup over that history — not fine-tuning.
Chatwootintegrated into the live inbox.
Handoffto a person when the assistant is unsure.
A support assistant that retrieves from three years of real conversations and lives in Chatwoot, with a defined route to a person when the question is outside that history.
A knowledge-backed support assistant — designed, connected and reviewed as in-house work at a Lithuanian conference organiser, documented here as a Zyndix project example.
Sitting on years of support history?
Bring a real question and the source that should answer it. Start with a free 30-minute call.