Qdrant RAG faces semi-structured edge cases

The store is not the bottleneck; the representation is. Semi-structured exports usually need a normalization layer that preserves hierarchy, identifiers, and parent-child context before embeddings ever enter the picture.

• –Chunk by object or record boundaries first, then add breadcrumb metadata so retrieval can rebuild context from nested fields instead of flattening everything into prose windows.
• –Use hybrid retrieval plus reranking for technical exports; dense search alone tends to miss exact names, config keys, and opaque IDs that matter in iPaaS payloads.
• –Low-confidence gating should combine score thresholds, top-1/top-2 score gaps, and retriever agreement; if signals are weak, abstain or ask a clarifying question instead of forcing an answer.
• –Build eval sets around rare schemas and edge-case transforms, because benchmark curves on common cases will look fine while long-tail queries keep failing in production.
• –The admin-only model switcher is useful for benchmarking, but it will not compensate for poor upstream parsing or missing metadata.