IskayPet replaced a manual SharePoint search process with an Amazon Bedrock Managed Knowledge Base that indexes the same document set and powers a multilingual employee assistant. Engineers care because the change eliminates the need to provision and maintain vector databases, embedding pipelines, and custom ingestion code while still delivering citation‑rich answers.
Why the Managed Knowledge Base mattered
The company needed a system that could answer natural‑language questions across four languages, respect document‑level permissions, and scale as new stores and policies were added. Bedrock Managed Knowledge Bases provided a fully managed RAG pipeline that handles ingestion, multimodal parsing, vector storage, embedding generation, and retrieval. The service also offers a native SharePoint connector, so existing repositories could be used without migration.
Architecture and implementation
The assistant follows a straightforward flow:
- An employee submits a question through the chat UI.
- The backend forwards the query and the employee’s identity to the Bedrock knowledge base.
- Bedrock runs a hybrid search that combines semantic similarity with BM25 keyword matching, then applies a managed reranking step.
- The top passages, each tagged with source metadata, are returned to the application.
- The application invokes Claude Haiku 4.5 (via Kiro) to generate a grounded answer that includes citations in the employee’s language.
Key implementation points include:
- SharePoint connector: Directly pulls documents from the existing SharePoint environment; advanced indexing also extracts text from embedded images.
- Hybrid search: Guarantees retrieval of exact terms such as product codes that pure semantic search might miss.
- Document‑level access control: Retrieval respects the permissions defined in SharePoint, limiting results to documents the user can already read.
- Model flexibility: The same knowledge base can switch underlying models without code changes.
- Built‑in evaluation: Quality can be measured at each ingestion or query step using Bedrock’s evaluation tools.
Operational and security implications
From an operations perspective, the team no longer runs separate vector store clusters or custom embedding services, reducing surface area for scaling and failure. Billing becomes usage‑based, tied to storage volume and retrieval count, simplifying cost tracking.
Security considerations shift to the managed service’s permission model. Because Bedrock enforces SharePoint ACLs at query time, engineers must ensure that identity information is correctly propagated from the front‑end to the Bedrock API. Any mis‑alignment could expose documents to unauthorized users. The managed reranking and evaluation features are provided at no extra charge, but they still rely on the same underlying data, so data‑retention policies for SharePoint apply unchanged.
Related CloudNinjas coverage: AWS.
What This Means For Practitioners
Adopting Bedrock Managed Knowledge Bases lets you replace custom RAG infrastructure with a service that handles ingestion, indexing, and retrieval while preserving existing document permissions. Evaluate the connector’s coverage for your source systems, verify that identity propagation meets your access‑control requirements, and monitor retrieval costs as usage grows. Future work may include testing additional foundation models within the same knowledge base or extending the pattern to other SaaS document stores.
