Indos.pl customer assistant

Visitors to Indos.pl often need quick answers about financial products before they are ready to contact an advisor. The assistant was created to provide useful, source-grounded responses directly on the website while keeping business knowledge under the control of the content team.

I designed and implemented the complete assistant, including the conversational flow, retrieval pipeline, Payload CMS integration, knowledge synchronisation, rate limiting, response rendering, and lead-notification workflow. I also made the system prompt editable from the CMS.

Answers needed to stay grounded in approved company materials rather than rely on the model's general knowledge. The knowledge base had to remain synchronised with CMS content, and the public endpoint required abuse protection. The solution also needed to recognise contact details without interrupting the conversation.

The RAG flow has three stages. First, a model analyses the latest message together with the conversation and prepares a focused retrieval query. Next, Gemini Enterprise Agent Platform searches the Discovery Engine knowledge base. Finally, the selected entries are passed to the response model, which writes a customer-facing answer. A dedicated integration synchronises Payload CMS content with Google documents, Upstash Redis limits traffic, and react-markdown safely presents structured responses.

The assistant now handles common product questions on the main website and gives the content team control over both knowledge and behaviour. When a visitor shares an email address or phone number, the system detects the lead and notifies an employee by email, turning a support interaction into an actionable follow-up without adding friction to the chat.

The quality of a production assistant depends as much on content operations as on model selection. Separating intent analysis, retrieval, and answer generation made the flow easier to observe and improve, while CMS-driven prompts and synchronisation gave the business a practical way to keep responses current.