Zenova Chatbot.

Key challenges

The hard parts, and how we solved them.

• Ensuring response accuracy across diverse document types

  →The knowledge base spans PDFs, Word documents, and Excel sheets — each with different structure and formatting. We built a unified ingestion pipeline that processes all three formats into clean, chunked embeddings, ensuring the retrieval layer can surface relevant content regardless of source format.

• Preventing hallucination on product-critical information

  →For a company dealing with fire suppression products, fabricated specifications or incorrect safety data would be unacceptable. The system prompt was engineered to constrain the AI strictly to retrieved content, with an explicit fallback instruction to acknowledge when a query cannot be answered from available documents rather than improvise.

• Keeping admin in control without developer dependency

  →The admin needed to update the knowledge base as new products and certifications came to market — without raising a ticket to SprintX every time. The OneDrive sync integration and manual upload/delete functionality in the admin panel gave them full autonomy over the knowledge base post-launch.

• Data privacy and security requirements

  →The client required that no company data be stored or retained by third parties. The architecture was designed so all document storage and vector embeddings remain within the client's own database, with AI API calls being the only external touchpoint — and those transmit only the minimum necessary content per query.