The assistant on nevioxdigital.com answers questions about our services, process and articles from content we have actually published, not from the general knowledge of the model. It is the same RAG approach we build for clients, tested on our own site first.
The goal: an answer right away, nothing made up
A visitor with a question at ten in the evening does not want to wait for an email. But a chatbot that confidently invents a price or a deadline is worse than none. We needed an assistant that answers right away, only from what is on the site, and that can say when it does not know.
What we built
Answers from our own content
Pages and articles are split into short passages and stored in a vector database (Pinecone). When someone asks a question, the system first finds the most relevant passages, and only then does the model write the answer, based on those passages and the latest messages in the conversation.
Three languages without manual translation
The search looks at content in the page language first. If there are not enough good passages there, it also looks at the other two languages. The answer comes in the language the visitor asked in, not necessarily the language of the page.
When it does not know, it says so
If the answer is not in the content, the assistant says so and points to the contact page. It does not invent prices, timelines, client names or guarantees; for those it always refers the visitor to us.
A new article goes straight into the index
When we publish a new article, it is added to the index automatically, so the assistant knows about it without manual work. The same index also powers related-article suggestions on our site.
Costs under control
Message length, how much of the conversation goes to the model and the number of requests per visitor are all limited. A single conversation cannot create an unexpected bill.
Results
- The assistant runs across the whole of nevioxdigital.com, in Croatian, English and German.
- Answers rest on published content, not on the general knowledge of the model.
- When there is no answer, the assistant says so instead of guessing.
- New articles enter the index automatically when they are published.
