Post · 01

Retrieval-first AI

Positioning thesis for retrieval-first AI.

Most AI products inject context at inference time and call it personalization. This thesis argues the primitive is wrong.

Retrieval-first means the system earns the right to reason by finding what is actually relevant first. Not injecting, retrieving.

Part 1 of 5 in the Declared First series, with Declare Your Frame (/journal/declare-your-frame), Pattern Weaver (/journal/pattern-weaver), Àṣà (/atelier/asa), and What I-O Psychology knows about human data that ML doesn't (/journal/io-psychology-and-human-data).

Reading as
Hiring team

Why a PM wrote an AI architecture thesis

Because product decisions are architecture decisions. If you pick the wrong primitive at the system level, no amount of prompt engineering saves you. This piece shows how I think before I build.

  • Systems-level product thinking
  • AI infrastructure literacy
  • Thesis-driven decision making
What you'll see: The argument in plain English, then the architecture implications for product teams.

RAG is not retrieval-first

RAG bolts retrieval onto generation. Retrieval-first inverts the dependency: retrieval defines what gets reasoned over. The distinction matters more than it sounds.

  • Retrieval as a first-class primitive
  • Why embedding similarity is not enough
  • Architecture patterns that follow from the thesis
What you'll see: Technical depth without the jargon wall.

AI that finds before it speaks

Most AI tools guess what you need. This explores what it looks like when a system is required to find relevant context first, before it is allowed to say anything.

  • Why "relevant" is a harder problem than it looks
  • What this means for tools you already use

Tell me what angle you are coming from

This piece lives at the intersection of AI product design and systems architecture. Happy to point you to the right entry point.