How IskayPet Cut Answer Time from Minutes to Seconds with Amazon Bedrock Managed Knowledge Bases
According to the AWS blog, IskayPet built a corporate knowledge assistant on Amazon Bedrock Managed Knowledge Bases and saw answer time drop from 3–5 minutes to about 7 seconds. The change matters because frontline staff spend a lot of time hunting for policies that live in scattered SharePoint files across four languages.
The problem: scattered docs and slow answers
IskayPet operates more than 330 stores in Spain, Portugal and France and supports 4,500 employees. Operational policies – return rules, veterinary protocols, onboarding steps – lived in hundreds of SharePoint documents. Employees had to open many sites, search manually, and often got outdated or wrong information. The company needed a system that could answer natural‑language questions, respect language and country boundaries, and reuse the existing SharePoint permissions without a massive migration.
How the managed service works
The Bedrock Managed Knowledge Base takes care of the typical RAG stack:
- Ingestion – a native SharePoint connector pulls documents, extracts text, and even parses visual content from images inside PDFs, DOCX and PPTX files.
- Chunking and embedding – the service automatically splits texts into chunks and generates vector embeddings, so no separate pipeline is required.
- Hybrid search – a combination of semantic (meaning‑based) search and BM25 keyword search runs on every query. This recovers exact terms like product codes that pure semantic search can miss.
- Reranking – the retrieved passages are reordered by relevance at no extra charge.
- Access control – the user’s SharePoint permissions are checked at query time, so each employee only sees passages they are already allowed to open.
The assistant calls the knowledge base, receives up to 25 passages, discards duplicates and short fragments, and injects up to 15 passages into a prompt for Claude Haiku 4.5. The model is instructed to cite sources and to say when the answer is not present in the retrieved content.
Managed vs. self‑managed RAG – a quick comparison
| Aspect | Managed Knowledge Base (Bedrock) | Self‑managed RAG stack |
|---|---|---|
| Vector storage | Amazon‑hosted, pay‑as‑you‑go | You provision and operate a vector DB (e.g., Pinecone, Milvus) |
| Embedding generation | Included in ingestion cost | Separate compute (often separate API calls) |
| Document parsing | Built‑in SharePoint connector, visual indexing | Custom pipelines for PDF, image OCR, etc. |
| Hybrid search | Semantic + BM25 provided, no extra charge | You must combine services or build your own ranking layer |
| Access control | Document‑level SharePoint permissions enforced automatically | You need to sync permissions and enforce them in your app |
| Operational overhead | No vector DB ops, no cron jobs for sync | Ongoing maintenance of ingestion jobs, DB scaling, monitoring |
| Cost focus | Storage, retrieval API calls, generation tokens | Additional compute, DB storage, network egress |
The table shows that the managed service removes a large chunk of engineering work – the vector store, embedding pipeline and permission syncing – and shifts most cost to storage and token usage.
What the trade‑off actually looks like
The headline benefit is speed: the average answer now arrives in about 7 seconds instead of the 3–5 minutes employees spent clicking through SharePoint. That speed gain comes from two things. First, the hybrid search finds the right passages quickly because it can fall back to exact keyword matches when a semantic match would miss a product code. Second, the knowledge base is already indexed, so there is no latency from a separate embedding step at query time.
The hidden cost is token consumption. Each query injects up to 15 passages, and those passages can be several hundred tokens each. Input tokens dominate the monthly bill, especially as usage scales. IskayPet is already looking at prompt‑caching for the static part of the prompt and batch inference for non‑interactive workloads to keep the bill in check.
Another subtle point is flexibility. Because retrieval and generation are decoupled, IskayPet can swap Claude Haiku 4.5 for a more capable model later without touching the ingestion layer. That flexibility is a vendor claim that the blog confirms – the architecture truly allows a model swap with minimal code changes.
Who should consider a managed knowledge base?
- Teams with existing SharePoint or similar document stores – the native connector avoids a costly data migration.
- Organizations that lack a dedicated ML ops crew – the service eliminates the need to run a vector database and manage embedding pipelines.
- Multi‑language, multi‑region groups – a single Bedrock instance can serve four languages and three countries, applying per‑assistant filters to keep knowledge isolated.
- Businesses that need strict access control – the built‑in document‑level permission check means you do not have to re‑implement SharePoint ACLs.
Conversely, a self‑managed stack may still make sense for companies that already run their own vector store at scale, need fine‑grained cost control on embedding compute, or have highly custom preprocessing requirements that the managed service does not expose.
Practical next steps you can take today
- Audit your internal docs – list the repositories (SharePoint, Confluence, file shares) that hold the most frequently asked‑about policies.
- Create a free‑tier Bedrock account and enable the Managed Knowledge Base preview.
- Connect a single SharePoint library using the native connector and enable visual indexing.
- Run a hybrid search test: ask a question that contains a product code and verify that the returned passages include the exact code.
- Measure latency and token usage on the first 10 queries; compare it to the time your team currently spends searching manually.
If the latency improvement is noticeable and the token cost stays within your budget, you can roll the assistant out to a pilot group, add more libraries, and duplicate the setup for other languages or regions.
