AI Chat Over Your Filings, With a Citation on Every Number
Rexfin's chat answers finance questions from your filings and cites every figure, or it refuses. No number ever appears in the prose without a source atom behind it.
By The Rexfin team
You type “what was gross margin in FY2024 and how did it move against FY2023?” into a chat box, and something answers in a normal sentence with a number in it. That is the easy part to build. The hard part is making sure the number is real, and that you can prove it in three seconds if someone in the room doubts it. That is the part most finance chat tools skip, and it is the part Rexfin’s chat is built around.
Two jobs, kept apart
Rexfin’s answer engine splits the work into a step that reasons and a step that writes. First, a deterministic investigator retrieves the specific facts your question needs from the atom graph built during ingestion: the individual numbers, their statement lines, their pages. That evidence gets assembled into a fixed block. Only then does a language model write the sentence, and it is only allowed to see that evidence block. It never sees the raw filing text and it has no memory of other numbers to draw on.
Before the answer ships, a separate check reads the prose the model wrote and looks for numbers. Every one has to trace back to a cited fact in the evidence block. If the model writes a figure that isn’t backed by a citation (rounded differently, invented, or pulled from something outside the evidence), the answer is refused rather than sent. That refusal is not a soft warning. It is the difference between “I don’t have that in the provided filings” and a wrong number dressed up as a fact. You can go read how that trust boundary is enforced in more detail in cite-or-refuse answers.
Follow-ups don’t get to cheat
Chat only feels like chat if you can ask a follow-up. “And for 2023?” only makes sense if the system remembers what “and” refers to. Rexfin’s threads handle that by carrying forward the phrasing of a prior question, a canonical predicate like “gross margin”, never the prior answer, never its citations, never a number. The current question still runs the exact same retrieve-then-cite pipeline from scratch. A prior turn cannot smuggle an unverified figure into a later answer, and if the follow-up resolves to something with no supporting evidence, it refuses just like a first question would. A refusal upstream doesn’t get rescued by a vague follow-up either: it stays a refusal.
The answers you see also carry document and entity scoping. If you ask about “revenue” on a filing with a segment breakdown, the engine won’t hand you a segment’s number dressed as the consolidated total, and it won’t reach across an unrelated filing to answer your question. Where that number lives, and why the engine trusts it enough to cite it, is covered in how the trust chain ranks evidence.
Repeat questions come back instantly
Asking the same question twice (you, or a colleague, on the same workspace) doesn’t re-run the model. Rexfin caches the answer, keyed partly to a fingerprint of every document currently visible in that workspace. The moment anything in the document set changes (a new upload, a document flagged, a filing superseded), the fingerprint shifts and the cache entry can no longer match. You get a fast repeat answer when nothing has changed, and a fresh recomputation the instant something has, with no manual cache-busting required. A cached answer is marked as such in the interface, so nobody mistakes speed for staleness.
What this doesn’t do
This is chat over evidence you already have, not a forecasting tool. If you want to ask “what if” about a driver rather than a historical figure, that’s a related but different capability: see natural-language what-if. And if you want the same verified answer phrased differently for a board versus an analyst, that’s a presentation layer on top of this same engine, described in answer voices.
The honest limit worth naming: the engine can only cite what your filings actually disclose. If the number genuinely isn’t in the documents, the right answer is a refusal, not a plausible guess, and that is exactly the behavior this is built to guarantee, every time, not just on the questions someone happened to test.
For the wider pillar this sits in, start at Rexfin product tour.
Part of Rexfin Product Tour: Every Number Traceable