Not the Whole Filing: Why Every Rexfin Citation Points to One Page

Ask most finance tools where a number came from and you get a document name. “Annual Report 2024.” Sometimes, if you’re lucky, a page number typed in by whoever built the model, already stale the moment the filing gets re-paginated. That’s a citation to a folder, not to evidence. It tells you where to start looking, not what you’d find if you looked.

Rexfin treats “page 42” as a starting point, not an answer. Every citation, whether it’s a cell in the statement grid or a line in an AI chat answer, resolves to one page and to a specific location on that page. That precision isn’t cosmetic. It’s built by two separate pieces of infrastructure that have to agree before a citation is allowed to exist at all.

The producer side: one page, one stable artifact

Before a filing can be cited from, every page in it is turned into its own canonical artifact: a stable, page-level rendering that comes out byte-identical no matter how many times the document is re-processed. That stability matters more than it sounds like it should: if re-running extraction on the same PDF produced a slightly different page each time, a citation captured last week could point at a page that no longer matches what a reviewer sees today.

Sitting alongside those canonical pages is a locator index. Its job is narrow and specific: given a document and a page number, resolve to the canonical page artifact for that page, plus the exact address on it where a given piece of evidence sits. That’s the difference between “this number is somewhere on page 42” and “this number is at this position on this rendering of page 42”: a distinction that only matters once you actually try to show someone the page, which is exactly what happens in the provenance drawer when a cited cell is clicked. The drawer doesn’t guess where on the page a figure lives; it asks the locator index, which already worked that out from the page’s own geometry, reconstructing rows and anchoring on column headers rather than trusting that a figure will always sit in the same place twice.

The consumer side: an answer engine that can’t cite loosely

Statement cells are one thing. AI chat answers are harder to pin down, because they draw on narrative disclosures and notes (the material outside the three face statements), not just numbers sitting in a grid. That’s what the cited chunk corpus behind chat exists for: retrieval records, each one keyed to a citation that names a document and a page, covering cover notes and narrative disclosure alongside the classification and summary information the answer engine needs to reason over.

The part that keeps this honest is an emit-time check: every one of those chunks has its citation checked against the canonical page index before it’s allowed into the corpus at all. A chunk whose citation doesn’t resolve to an actual page isn’t a citation with a small error in it: it’s treated as a defect and kept out. An AI answer built on this corpus can only ever cite a page that’s known to exist and known to be reachable, because nothing without that guarantee made it into the pool of things the answer engine is allowed to quote.

Why the document isn’t the unit that matters

A citation that names a filing but not a page asks you to trust the summary. A citation that names a page but not a location asks you to trust that you’ll find the right number among however many are printed there. Neither is verification: both are still, at bottom, someone’s word. Rexfin’s citations are built backward from the smallest unit that can actually be checked: a specific spot on a specific, stable rendering of a specific page. Everything downstream (the ringed figure in the drawer, the resolvable line in a chat answer, the grid cell you started from) is built on top of that unit, not around it.

See how this fits with the rest of the trust chain at the product tour hub.