isanna shop

What it is

A decision written down as a research — the question, the buyer’s own context, and how often it should be redone. On each cycle an agent goes and does it: it chooses where to look, reads, and leaves behind products, prices, specifications and claims, each one carrying the page it came from.

Keeping the original page pays for itself. When a shop’s price turned out to be missing from the extracted text, the number was already inside the stored bytes — in the page’s own structured data — so the fix was a parser change replayed over documents fetched hours earlier, without going back to the shop.

How it works

The research

A question, a cadence, and the context that makes the answer the buyer’s own — what the machine is for, what the budget is, what has already been ruled out. The context is read at the start of every cycle, so a preference written once shapes every run after it.

The cycle

A scheduler wakes, sees which research is due, and hands it to an agent — Codex or Claude — with a tool surface scoped to that run. The agent picks its own sources: there is no approved list, and a domain earns its place by yielding facts.

The graph

SQLite — products, offers, specifications and claims — with the page behind each fact stored as it was fetched. One row per configuration, so 96 GB and 256 GB of the same machine are two products, and prices at home and abroad are separate series that are never converted into each other.

Asking

Questions are put to the archive in plain language, over the graph and the text of the stored pages, and the answer cites what it rests on. When the archive does not know, that is the answer.