RAG Is Not a Database
Retrieval-augmented generation is usually built as though a vector store were a database with fuzzy matching. It is a ranking system, and treating it as one changes almost every design decision.
6 min read
Writing
Long-form pieces on software engineering, distributed systems, cloud architecture and AI. Roughly one substantial article a week.
Retrieval-augmented generation is usually built as though a vector store were a database with fuzzy matching. It is a ranking system, and treating it as one changes almost every design decision.
6 min read
A distributed lock cannot promise mutual exclusion, because it cannot control the process holding it. What it can do is still useful — provided you know which of the two problems you have.
6 min read
Kafka gives you ordering per partition. Almost every ordering bug in production comes from a gap between that guarantee and the one people believed they had.
6 min read
Every abstraction is a bet about what you will never need to know. Understanding the bet is most of what separates engineers who can debug a system from engineers who can only operate one.
6 min read
Subscribe via RSS, or see every topic.