Explanation

Why Option C is correct:

1. Amazon Bedrock Knowledge Bases with RAG addresses the core issue of recommending products not in the catalog:

   • RAG retrieves relevant information from the product catalog before generating recommendations
   • Ensures recommendations are grounded in actual available products
   • Knowledge Bases automatically manage the retrieval process and vector embeddings

2. PerformanceConfigLatency parameter addresses the response time issue:

   • Setting it to "optimized" prioritizes lower latency
   • This helps reduce the long response times reported by customers

Why other options are incorrect:

Option A:

• Amazon Bedrock Guardrails focus on safety and content filtering, not ensuring factual accuracy about product availability
• Automated reasoning checks don't solve the problem of hallucinated products
• Provisioned throughput helps with throughput but doesn't address relevance or accuracy issues

Option B:

• Prompt engineering alone cannot reliably restrict model responses to only existing products
• LLMs can still hallucinate products even with careful prompting
• Streaming inference reduces perceived latency but doesn't solve the accuracy problem

Option D:

• While OpenSearch can store catalog data and validation can check recommendations, this is a manual approach
• RAG in Bedrock Knowledge Bases provides a more integrated, automated solution
• Response caching helps with repeated queries but most interactions are unique (as stated in the problem)

Key takeaway: The combination of RAG (to ground responses in actual catalog data) and latency optimization (to improve response times) directly addresses both customer complaints mentioned in the scenario.