Company materials become a reviewed knowledge base. Customers and employees get clear answers in the right tone — on a standalone page or directly on the company website.
Built and operated by UNECT — running on this site and for companies in retail and services
Customers repeatedly ask about opening hours, prices, services and the next step. The answers often already exist somewhere on the website, but conventional navigation and site search still leave people hunting — while staff answer the same e-mails again and again.
The assistant starts with the company's own materials. It discovers pages through the sitemap and internal links, prioritizes useful sections such as the offer, FAQ and pricing, filters low-value pages, then scrapes and cleans the selected content. The resulting sources can be reviewed before the assistant goes live.
When a question arrives, the engine ranks the available material and supplies only relevant context to the language model. Depending on the volume, it uses either full documents or heading-based RAG chunks. Company instructions set the language and tone; when the source does not contain an answer, the assistant is instructed to say so rather than improvise.
A branded demo is the first step. After the content and answer style are approved, the same assistant can run as a standalone experience or through a widget on the company's own site. The assistant in the bottom-right corner of this page uses the same underlying engine.
A scripted example of the product experience. Source labels make the important distinction visible: the assistant answers from company material, not from a generic guess. The live UNECT assistant in the corner runs on the same engine.
AI assistant
Do you deliver dining tables outside the city?
Yes. Standard delivery covers addresses within 120 km, and longer-distance delivery can be quoted individually. Assembly can be added at checkout.
From: Delivery & assembly pageAnd what does assembly cost for a table?
Assembly for a dining table is listed at €45. If your order includes several items, the team confirms the combined price before delivery.
Source: /pricingSuggested questions
How the assistant is prepared
What happens at answer time
A production assistant needs more than a prompt. Retrieval limits, privacy controls and explicit failure paths keep the experience predictable when providers or content fall short.
The primary route is retried, then a separate fallback model can take over. Every provider call has a 35-second timeout.
If every model attempt fails, the engine still returns a deterministic excerpt from the retrieved knowledge base and directs the visitor to the company.
Provider requests explicitly deny data collection and require zero data retention; configured providers are selected for low latency within those rules.
Full-context mode ranks documents within a 60,000-character budget; RAG selects up to six chunks within 14,000 characters.
Assistants have active flags and optional expiry dates. Demonstration pages are noindexed and unavailable after deactivation or expiry.
Successes and failures record the model route, provider, token counts and reported cost, while product events show how visitors use the assistant.
Technology: Next.js · TypeScript · OpenRouter multi-model routing · custom RAG · Cheerio scraping · JSON store
The current core answers grounded questions reliably. These are practical extensions for production deployments.
Connect booking, enquiry and qualification forms so a useful answer can lead directly to the next step.
Add governed ingestion for policies, manuals and other internal documents beyond public website pages.
Reuse one reviewed knowledge base across the website widget, standalone assistant and selected messaging channels.
Use conversation and event analytics to identify missing content, refine suggested questions and improve weak answers.
The first step is a focused demo built from your real website content. You can review the sources, test genuine customer questions and decide what belongs in production.
Discuss your assistant