firehose> #llmops

Query-Shaped Storage

The principle underneath every level of Retrieval Maturity Levels: how data will be accessed and recalled determines how you should put it in. You work backwards from the question — reverse-engineer the schema, folder layout, and retrieval style from how you intend to ask later, not from how the data happens to arrive. The analogy: you know a basketball hoop is round, so you'd never design the ball as a giant square — start with the end in mind. In practice this is what makes the level choice per-folder (Semantic Retrieval vs a plain markdown file vs Knowledge Graph Retrieval) a design decision rather than a default: ask "what type of questions will I ask of this data?" first, then pick the storage that answers them cheaply and accurately.

This is why the same failure recurs when it's ignored — e.g. chunking a meeting transcript into vectors when the question you'll actually ask is "summarize the whole meeting" (chunk retrieval can't see the whole document). The concrete tactic Nate recommends: ask the agent — "here's my data, here's how I want to use it; markdown files or semantic search?" — and let the query shape the decision.

Claims

Linked from