Knowledge search with cited sources: cite or decline

The most dangerous answer of a knowledge search is not the wrong one. It is the wrong one that sounds convincing.

An example: someone asks for the notice period in the framework contract. The search finds no matching passage and answers anyway, fluently and with a specific figure. Language models phrase things plausibly even when they lack the basis. Nobody notices until it gets expensive.

1. 1Every answer names its source: file, page and line range.
2. 2If the search finds no passage that supports the answer, it says so, without exception.
3. 3An answer without evidence is never shown.

That does not make the search worse. It makes it checkable: whoever reads an answer can look up what it rests on.

The basis is a method known as retrieval-augmented generation: first, matching passages are searched in your documents, then a language model words the answer on their basis. Turning that into an answer with evidence takes three things.

1. 1Provenance on ingestion: every passage is stored with file, page and line range. A passage without provenance is never stored.
2. 2Permissions before the search: only passages the person asking may see are searched.
3. 3A check before output: if none of the passages found supports the answer, the system declines instead of guessing.

Such a search can be measured. You ask it questions whose answer is in the corpus and questions whose answer is not. Both are measured: how often an answer is correctly backed, and how often the search rightly declines.

The safeguards themselves need tests too. In the open-source project cite-or-decline, each of the 36 guards is removed on purpose, one at a time, to show that its test goes red.

A search that sometimes says it finds nothing in the documents seems less helpful at first. In fact it is the one where you know where you stand: every answer brings its evidence, and every gap is named. A search people trust is a search people use.