Ten portals, one question box, and an answer that cites its source

The hard part was never the model. It is that the sources disagree about everything except the facts. One portal identifies a company by its fiscal code, another by its registration number, a third by a name typed by hand with the legal form in a different place each time. A tender is a page on one site, a structured record on another and a PDF attachment on a third. Before anything can be asked, all of it has to become one shape.

So every source implements the same contract and nothing downstream knows which kind it was. Decide whether this adapter handles the source, fetch a page, parse the items out of it, normalise them into a fixed set of fields — title, source URL, published date, deadline, value, currency, classification code — and report whether another page exists. A portal that has to be read as HTML and a feed that arrives structured produce identical records, which is what makes the next source a class rather than a project.

Ingestion that runs once is a demo. Each source carries its own expected cadence and is evaluated against it, producing a health verdict per source rather than one number for the whole system. That distinction matters more than it sounds: a platform aggregating dozens of registries will always have one that changed its markup last night, and the useful question is never “is the data fresh” but “which of these is stale, and since when”.

The links between sources are the product. A company record is not a page about a company; it is the join point for everything the other registries say about the same entity — the tenders it bid on, the contracts it won, the court files it appears in, the filings it made. That join is why the platform answers a question that no individual portal can: not “what does this registry hold about this company” but “what is true about this company”, and it is computed rather than curated.

Search runs eight corpora in one pass — companies, tenders, awards, procurement plans, EU notices, direct purchases, legal documents and court files — through a single multi-index query rather than eight sequential ones. The corpora stay separate because their fields are genuinely different, and the result set is merged after the fact with the type carried through, so a reader can see at a glance which kind of thing each answer is.

The assistants are where a platform like this usually starts lying, and the design is built around not letting it. The retriever is a contract with two drivers behind it — one comparing whole documents, one comparing passages — and what it returns is not text but a structure: the passage, the issuing authority, the URL, and where available the heading path and the article number it came from. The citation is a field on the result. An answer can therefore be checked against the thing it was drawn from, which is the only version of an AI answer that is worth anything in a professional context.

The same restraint runs through the query interpreter. A question typed in ordinary language goes to a model, which extracts slots — a sector, a county, a fiscal status — and stops there. The model never executes anything: the extracted slots go through a deterministic mapper into validated filters, and those filters run the same code path the manual builder runs. The interface then states, in one plain sentence built without a second model call, exactly which filters are about to execute, so an interpretation that got it wrong is visible and editable before it returns anything. Language is treated as an input method, not as an authority.