architecture

September 3, 2026

Machines Should Retrieve, Models Should Reason

Many search and matching workloads don't need an LLM or a vector database. Use the cheapest computational primitive capable of solving the problem.