Implement advanced RAG with Azure AI Search

Why it matters

An agent’s answer is only as good as the evidence it retrieves. This exercise turns chunking, embedding and ranking choices into measured trade-offs instead of assumptions.

What to practise

Generate fixed-overlap and structural parent-child chunks, upload both to three category indexes in Azure AI Search with managed identity, and give each chunk a content-only and a content-aware vector. Route each query by intent, then run vector, hybrid and semantic searches with SearchClient and VectorizedQuery. Compare MRR and latency by strategy, mode and embedding profile against chunk count. The exercise retrieves evidence only; it builds no agent.

How it fits

The chunk strategies are 2.2.8, the two embedding profiles 2.2.9, and the vector, hybrid and semantic stages 2.2.10. Routing to category indexes with source-specific citations is 2.2.11. For 3.1, note that MRR scores the rank of the expected parent, not answer quality, which is 3.1.6’s separation.