IskayPet’s Bedrock knowledge assistant makes permissions part of retrieval

The problem is finding the authorised procedure

IskayPet describes a business spanning more than 330 stores, over 4,500 employees and operations in Spain, Portugal and France. Its materials cover retail, veterinary care, procurement and onboarding in Spanish, Portuguese, French and English. These figures describe the organisation; they do not establish that every employee used the assistant or that every customer received AI support.

The practical problem was scattered SharePoint documentation. An employee might need a country-specific return policy while another needs a veterinary procedure. A fluent answer from the wrong country or business area can still be operationally wrong. The relevant unit of success is therefore an authorised, current answer with a usable source, not simply a generated sentence.

Our analysis is that this framing explains the architecture better than the language count alone. Supporting four languages matters, but so does preserving which documents belong to which operational context. The source presents a customer implementation, not a newly universal entitlement for Bedrock users.

Separate knowledge domains before generating answers

The team considered one shared knowledge base with country and area tags plus a filter on each query. It instead chose a dedicated knowledge base per assistant. Each assistant points to relevant SharePoint sites and can narrow ingestion to specific document libraries or folders. A new area or country can receive another assistant without rebuilding the whole system.

The native connector reuses the existing repository and synchronises new or changed documents incrementally. With advanced indexing enabled, the article says visual content inside PDF, DOCX and PPTX files, as well as standalone images, becomes searchable. This is a reported ingestion capability; we did not test whether every scanned document is parsed correctly.

Separate indexes make the intended scope easier to inspect and tune, but they do not replace authorisation. Nor does incremental synchronisation prove that every answer reflects the latest edit instantly: the article supplies no measured content-refresh interval. Its written architecture description is the evidence used here; the separately hosted diagram was not inspected.

Two permission checks serve different purposes

The first check determines which assistants a user can see. According to the case, visibility follows existing direct, group or folder permissions; if entitlement cannot be confirmed, the assistant is hidden. That limits the entry points presented to the employee.

The second check happens inside retrieval. The application passes the employee’s identity to the knowledge base, which returns passages only from documents that employee may read. Two employees can legitimately receive different answers to the same question. Hiding an assistant alone would not provide this document-level boundary.

The authors say access checks read SharePoint permissions, so changes can be reflected without rewriting the application. This remains an implementation claim, not our security certification. A useful acceptance test would use users with different permissions, revoke a document entitlement, and check both retrieved evidence and the final answer. The article does not report such a test suite or the timing of permission propagation.

Twenty-five retrieved passages are not fifteen free answers

Retrieval combines semantic matching with keyword search using BM25, then reranks the results. Semantic matching can recover related wording; keyword matching helps retain exact product codes and references. IskayPet retrieves 25 passages, removes duplicates and fragments too short to help, and supplies at most 15 to the generation prompt.

The counts describe two stages of one question. They are not a success rate or evidence that spending fell by 40%. Passage length, repeated context and generation length still affect the bill. AWS says managed parsing, embeddings and reranking add no separate charge in this arrangement; storage, retrieval calls and generation-model tokens remain cost components, with input tokens dominating this deployment.

The prompt requires citations and an admission when retrieved evidence lacks an answer. These instructions improve the intended behaviour but do not prove that hallucinations disappear. The production generation model is not named. Claude Haiku 4.5 is explicitly identified as the model used with Kiro to develop the frontend, architecture and deployment, so it should not silently become the asserted production answer model.

Guardrails are described as being added. Prompt caching and batch inference are under evaluation, with batch aimed at non-interactive work. They are future or ongoing measures in this account, not already demonstrated savings. Feedback from thumbs-up and thumbs-down usage informs evaluation, but the article gives no quantitative retrieval or answer-quality scores.

Read the seven-second result with its denominator

For roughly the first two months in production, the authors report that finding an operational answer fell from 3–5 minutes to about 7 seconds, alongside approximately 95% less time searching for procedures and policies. Those are customer-reported search-time observations. They do not measure a 95% reduction in total employee work, establish accuracy, or isolate the contribution of one model.

The arithmetic is broadly compatible with the claim: seven seconds compared with 180–300 seconds implies about 96–98% less elapsed time, before accounting for approximate timings and differing definitions. It would be misleading to replace the source’s approximate 95% with a more precise universal number. No sample size, task mix, baseline protocol or distribution of slow responses is provided.

The source expects around 100,000 monthly queries across areas and countries; that is not a measured production volume. Our practical conclusion is narrower: managed retrieval can reduce infrastructure work while leaving the organisation responsible for document quality, permissions, evaluation and spend. Before expanding a similar assistant, measure complete tasks, citation correctness, cross-country isolation and failed-access behaviour on an evaluation set that includes unsuccessful questions.