Industry Problem Solution Outcome
InsuranceDefault chunking broke on a mixed set of long policies and short FAQs Two data sources in one Knowledge Base, with hierarchical chunking for policies and semantic for FAQs Accuracy up roughly 25 points
Wealth management Retrieval worked, but the system prompt had no refusal path Rewrote the prompt with refusal logic, added a sufficiency check via InvokeModel Hallucination rate dropped to under 3%
Logistics SaaS No metadata filter, so every query searched the full document set Bedrock filter expressions limited each query by client ID and document category Monthly query costs cut by roughly 80%
Logistics SaaS One Knowledge Base per tenant hit Bedrock's 100-KB account limit at 80 customers Single Knowledge Base with tenant_id metadata filtering and IAM-level session scoping Onboarding ceiling removed, per-tenant costs eliminated
Legal Tech Reranking pulled for latency, complex queries returned shallow matches with bad citations Restored reranking via Cohere Rerank 3 on an async path Answer quality recovered