How IskayPet unified corporate knowledge with Amazon Bedrock Managed Knowledge Bases

IskayPet had the kind of knowledge problem that quietly wastes entire workdays. The answers were already there, buried in hundreds of SharePoint documents across four languages and three countries. Staff needed return policies, veterinary protocols, onboarding steps, procurement details — and they needed them fast, not after a scavenger hunt through folders.

The company built a corporate knowledge assistant on Amazon Bedrock Managed Knowledge Bases instead of running a self-managed RAG stack. That matters because the usual setup means owning vector storage, embeddings, ingestion jobs, and all the plumbing around them. Here, AWS handles the retrieval layer, while IskayPet gets a system that can pull answers from its existing SharePoint repositories without a migration project on the side.

The design is deliberately split by assistant. Each one points at specific SharePoint sites, and can be narrowed further to document libraries or folders, so knowledge stays isolated by area and country instead of turning into one giant mixed index. That choice also makes expansion simpler: a new area or country means adding an assistant and pointing it at the right content, not redesigning the whole thing.

Access control is part of the product, not an afterthought. The assistant only shows people the knowledge they’re already allowed to read, and it checks permissions again at retrieval time using the employee’s identity. So the same question can produce different answers for different staff, because the system is enforcing SharePoint permissions rather than inventing its own. It also cites source documents, which is a nice way of saying employees can verify the answer instead of trusting a black box that’s feeling confident today.

IskayPet also used Kiro with Claude Haiku 4.5 to build both the store-facing frontend and the AWS architecture. On the ingestion side, the SharePoint connector pulls documents in directly and can extract visual content from images inside PDFs, DOCX, and PPTX files, plus standalone images. In production, the company says time to find an operational answer dropped from 3–5 minutes to about 7 seconds, and the expected volume is roughly 100,000 queries a month across all business areas and countries.