Every grounded answer depends on a step that nobody sees. Before a model writes anything, something has to decide which passages out of a few thousand pages are the ones worth reading. That retrieval step sets the ceiling on everything downstream, because an answer can only be as good as the evidence placed in front of it, and a model given the wrong three paragraphs will produce a confident response about the wrong three paragraphs.
It is worth understanding how that selection works, partly because it explains a category of failure that looks like hallucination and is not. When an AI tool misses a fact that is plainly in your documents, the usual cause is not that the model could not read it. The passage was never retrieved, so the model never had it.
Keyword search matches the words you typed against the words on the page. It is exact, fast, and completely literal, which makes it excellent for the things professional documents are full of: a defined term, a clause number, a product name, a figure. Ask about “Section 7.2” and keyword matching finds Section 7.2 without ambiguity.
Its weakness is that documents rarely use your words. A contract discusses termination while you asked about ending the agreement, a report describes headcount reduction while you asked about layoffs, and literal matching finds nothing in either case even though the answer is sitting there. The passage exists, the vocabulary does not line up, and the search comes back empty.
Meaning-based search solves exactly that. It compares the sense of your question against the sense of each passage, so termination and ending the agreement land near each other regardless of shared words. The trade-off is that it is approximate by design, and approximation goes wrong in a specific way: it will happily return a passage about the wrong quarter, or the wrong subsidiary, because that passage is about the same kind of thing. Precise identifiers are where it is weakest, which is precisely where keyword search is strongest.
The failure modes are close to complementary, which is why the Evidence Base runs both. Keyword matching and meaning-based similarity search the same document set independently, and the two result sets are merged and ranked into a single ordered list of candidate passages before anything is passed to a model. A passage that both methods surface rises; a passage found by only one still gets its chance to be considered.
The practical effect is that a question mixing precise and fuzzy elements stops forcing a choice. “What did the Q3 report say about supplier concentration risk?” contains an exact token that keyword search handles and a concept that may be described five different ways across the document, which meaning search handles. Either method alone answers half the question well and half of it badly.
Retrieval quality is not an abstract concern, because it determines whether citations are worth anything. A citation is only useful if it points at the passage that actually supports the claim, and that passage has to have been retrieved in the first place. Weak retrieval produces answers cited to whatever was nearby, which look properly sourced until a reviewer opens one and finds it does not say what the answer claimed.
This is also why retrieval sits behind the source integrity step rather than in front of it. Finding the right passage in a stale document is not a success, it is a well-cited error, and no amount of retrieval sophistication compensates for a document set nobody checked. The order matters: audit the evidence, retrieve from it precisely, then let models work on what was found.
Qonera is the AI governance platform for professional teams, built around a structured review and approval workflow that starts with the Evidence Base: your documents indexed and integrity-checked, searched by keyword and by meaning together, with each answer carrying citations back to the passages behind it so a reviewer can check the claim rather than trust it. Most of what makes an AI answer defensible happens before the model produces a single word, in the unglamorous work of finding the right page.
Multi-model stress testing, Conflict Heatmap, tamper-evident audit trail, and structured sign-off, built for teams who need defensible AI output.