Issue 01 › Method

The least glamorous idea that fixes the most.

Retrieval is unfashionable, decades old, and solves more real problems than almost anything newer. Here is what it actually does.

We read the footnotes · 5 min read

Scattered document marks, a relevant few picked out in red and drawn toward a single block. — Only a handful of passages ever reach the model. Almost every disappointing system is failing here, at the choosing, rather than at the answering.

Fetch first, answer second

The idea is almost insultingly simple. Before the model replies, search your own material for the handful of passages most relevant to the question, put them in front of it, and tell it to answer from those.

Left of the line, choosing the passages. Right of it, answering from them. Almost every disappointing system is failing on the left and being blamed on the right.

That is it. No training, no fine-tuning, no model of your own.

Why it works so well

It converts the hardest question a model faces — what do I remember? — into the easiest one: what does this passage say? Models are genuinely excellent at the second. They are unreliable at the first, and no amount of scale has fixed that.

The rest is for subscribers.

It carries on in the same plain language, with the same drawn diagrams.

One week free, then ₹25