IskayPet’s seven-second knowledge search starts with employee permissions

A store employee looking up a returns rule needs the policy that applies to that store, in a language they can use, from material they are permitted to read. IskayPet has built its internal AI search around those boundaries. In an AWS customer account published on 22 October, the pet retailer reports reducing searches that took three to five minutes to answers in roughly seven seconds.[1]

The operation spans more than 330 stores and 4,500 employees in Spain, Portugal and France, with documents in four languages. Its assistants use Amazon Bedrock Managed Knowledge Bases to retrieve corporate information from sources including SharePoint. The account describes approximately two months of production use, rather than a long-term independent evaluation.[1]

Separate assistants, scoped information

Each assistant has a dedicated knowledge base, and access rules determine both which assistants an employee can see and which documents can be retrieved. That is a substantive design choice. If a model sees a restricted document before generating an answer, a polished response can already disclose information the employee should never have received.[1]

The retrieval pipeline combines keyword and semantic search. It initially gathers 25 results, removes duplicates and fragments that are too short, then supplies up to 15 to the model. This gives an answer relevant supporting material while controlling the volume of text sent for generation. The account identifies input tokens as a major cost driver.[1]

Seven seconds is an operating result; query volume is a forecast

IskayPet also estimates that employees spend approximately 95% less time searching. The expected volume of about 100,000 queries a month is prospective; the article does not establish that employees have already submitted that many.[1]

The company describes using Kiro and Claude Haiku 4.5 in development. That does not identify Haiku 4.5 as the model generating every production answer. It also describes further guardrails being added, so those plans should not be counted as controls already operating.[1]

The result is a concrete example of where internal search becomes useful: reducing the work between a routine question and the relevant policy. Its continuing value depends on maintaining the document scopes as employees, stores and rules change. Two people can ask the same question and legitimately need different answers.[1]

Report an error or suggest a correction