TALAMANA · THE AI LITERACY MAP FOR ARCHITECTURE AND DESIGN · Generative Mechanics · AGE 19—21 · POSITIONAL · HELD

Fine-tune or retrieve

The honest first move is retrieval, because training buries knowledge in the weights.

The dilemma

The studio wants the model to "know our work". There are two roads. Fine-tune, so the knowledge lives inside the weights. Or retrieve, so the knowledge stays in an archive you can edit and is fed in at the moment you ask. Which do you reach for first?

The choices

Train. You get a model that carries your manner without being told each time. It costs rights work and compute, and it freezes your archive at training day. Retrieve. You get an archive you can correct tomorrow, answers that can cite their source, and no training. It costs a retrieval pipeline that must be built well, and the model still sounds like itself. Or split the two: retrieve for facts and precedents, and train only for a visual signature that words cannot carry.

The consequence

Train first and you own a model that knows last year's studio, cannot say where an answer came from, and must be retrained when the archive changes. Retrieve first and you can show your working, fix your sources and switch base models freely. But a weak retriever hands the model the wrong document with full confidence. The split takes more thought, and it is usually the honest answer.

The case

A twelve-person practice wants "our standard details and our way of writing specs" in the model. The knowledge is documents: details, specifications, past reports. Retrieval serves it at once, with citations to the detail sheet. Six months later they want renders in the house visual language. Words do not carry it, and fifty cleared images do. That is the moment for a LoRA, and not before.

Our position

For most studios the honest first move is retrieval, not training. Knowledge that changes, that must be cited, or that lives in documents belongs in an archive you retrieve from. Training is reserved for what words cannot carry — a visual signature — and only after the zero-training paths are exhausted. We hold this as studio policy: a default order for a practice with limited time and rights, not a ranking of the techniques.

Why we hold it

Retrieval keeps the studio's knowledge editable, inspectable and portable across models. Training buries it in the weights. In our teaching, the studios that trained first took on upkeep they had not priced. The ones that retrieved first kept their citations and their options.

The strongest objection

Retrieval and fine-tuning are not first and last answers to one problem. They change different things. Retrieval supplies information at the moment you ask. Fine-tuning changes behaviour — format, domain habits, how the model selects tools, its manner, its style. Sometimes the right architecture is retrieval, sometimes fine-tuning, sometimes examples in the prompt, sometimes two together, and the published comparisons show task-dependent results rather than a universal order. So "retrieve first, train last" is defensible as a studio policy and not as a technically privileged sequence — and a card that presents policy as mechanism teaches the wrong kind of reason. A practice whose value is a way of reasoning rather than a set of documents may be delayed by the policy from the one thing that would work.

What would make us revise it

Measured evidence, on real studio tasks, that the default order is the wrong default: that a light fine-tune for behaviour plus retrieval for facts beats retrieval alone often enough that "retrieve first" costs studios more than it saves. Or training becoming cheap, rights-clean and reversible enough that the upkeep cost that justifies the policy disappears. Reviewed every edition.

Take it to crit

When a studio said it wanted the model to "know our work", did you reach for training out of habit, or weigh retrieval first and say why? Show the split.

How it works

Retrieval (RAG on this strand) places the archive's nearest pieces in context. Fine-tuning (FINE-TUNE) shifts the weights. The two can be combined, and many serious systems retrieve into a lightly tuned model. Sahil's essays on training data and on the studio stack set out the zero-training-first ladder this position rests on.

What this idea builds on

What this idea opens up

Sources

Open this idea on the map · The complete map · Logika · RBDS AI Lab, India · revised every edition.

Age grows from 11 at the centre to 22 at the edge, and six sectors show the learning strands. Tab into the map and the arrow keys step from idea to idea, following the links where there is one. Enter opens the idea under the cursor, and E reads out its links and the reason recorded on each. Press slash for Search, question mark for the full key list, and Escape to leave. Open Ideas for the complete readable list, including what each idea builds on and what it opens up.

LOGIKA · RBDS AI LAB INDIA
ON-RAMP · AGE 11 · FIRST ENCOUNTERS, NOT GATES — IDEAS · — DEPENDENCIES
DONE
OPENS NEXT
SOLID — STANDS ON