Retrieval Abstraction Layer

A retrieval abstraction layer is a thin evaluation harness your agency owns that sits between client applications and whichever RAG vendor supplies indexing and retrieval. The vendor handles parsing, chunking, and semantic search; your layer owns the golden question set, the accuracy scoring, and the swap decision. This matters because retrieval quality and pricing move independently across providers, and a client retainer signed against one vendor's benchmark can quietly degrade when that vendor changes its ranking model. Concretely, Perplexity's Photon retrieval engine cut p99 latency from 800ms to 65ms, a 12x improvement that makes latency a live evaluation criterion rather than a fixed assumption. Agencies that score retrieval on their own test set can move a client from Ragie to another context engine in days, and can price the migration as a billable accuracy audit instead of absorbing it as rework. The layer is the asset; the vendor is the commodity.