Hybrid Retrieval: How Rexfin Finds the Right Evidence Before the AI Answers

Ask a chat tool a question about a filing and, somewhere behind the fluent sentence it hands back, a search has to happen first: find the handful of passages that could actually contain the answer, out of every page of every document you’ve uploaded. Get that step wrong (miss the right line, or hand the model ten pages of near-misses) and no amount of careful prompting downstream fixes it. Rexfin treats that search as a problem worth solving properly, not a wrapper around whatever search box came bundled with the database.

Two ways to miss, so use both

Keyword search finds an exact term reliably (a line-item name, an account code, a specific phrase from a footnote), but it misses a question phrased differently than the filing phrases it. Semantic search catches the paraphrase, matching “cash position” against a filing that only ever says “liquid funds,” but on its own it can also surface something that merely sounds related without being the right fact. Used alone, either approach leaves gaps the other would have caught. Rexfin runs both against every filing: a lexical pass that matches on the actual words, and a dense pass that matches on meaning through vector search, and fuses the two ranked lists into one before anything moves further down the pipeline. Neither pass gets the final word by itself; the fusion is what decides what counts as strong evidence.

No model guessing at the question first

Some retrieval systems ask a language model to invent a hypothetical answer first, then search for passages that resemble that invention: a workaround for weak search, and a new place for the system to go confidently wrong before the real answer step has even started. Rexfin’s retrieval doesn’t take that detour. It searches directly on what was actually asked, against the actual filing text and its embeddings. One fewer model in the loop is one fewer place for an invented premise to quietly steer what gets treated as evidence.

Living next to the facts, not bolted on

The lexical index and the vector index both run inside the same database that holds the underlying facts, rather than depending on a separate search service you’d have to keep in sync with the filings as they’re ingested. That matters more than it sounds: search results only mean anything if they point back at facts. If the search index lives elsewhere and drifts out of step with what has actually been ingested, a retrieved passage can end up pointing at something you no longer trust, a subtly worse problem than being slow.

A shortlist the answerer can’t argue past

What comes out of retrieval isn’t a paragraph of prose or a summary: it’s a ranked set of specific chunks and atoms, each one traceable back to its exact page. That shortlist is the entire universe of evidence the answer engine is allowed to draw from for that question. It’s the reason the platform can commit to a hard rule further downstream: cite or refuse, never estimate. If retrieval didn’t surface the right passage, the answer engine has nothing to cite it with, and the honest response is a refusal rather than a fluent guess. The same shortlist is what makes it possible to show, next to every figure in chat, exactly which page it came from: retrieval is the step that makes that pointer available in the first place.

Checked, not just trusted

Retrieval sits upstream of the two things that matter most about an AI answer: whether the number is right, and whether it’s cited to something real. So it isn’t left to run untested once it’s built. It’s part of what the platform’s answer quality evaluation checks on every change: a change to retrieval has to clear the same graded test set as a change to the model or the prompts, and a version that quietly starts missing the right evidence gets caught there before it reaches an answer, not after.

Why this matters more than it looks

A search engine feels like plumbing: invisible when it works, and rarely the first thing anyone asks about. But it’s the layer that decides what the AI is even allowed to see before it opens its mouth. Get it wrong and the AI can be perfectly well-behaved about citing sources and still cite the wrong ones, because the right passage was never in the candidate set to begin with. Rexfin’s approach (lexical and semantic search combined, no query-rewriting model to trust blindly, running against the same store as the facts it’s searching) is built so that whatever the AI eventually says, it was looking at the right page when it said it. See the pillar overview for how retrieval fits with the rest of the platform, or book a demo to see it find the right footnote in a filing you bring.