How answers are picked

Your documents are split into passages of about 90 words, never across a heading. Each passage keeps its heading as context. Words are lower-cased, common words are dropped and simple endings are folded, so “recordings” and “recorded” meet at “record”.

Passages are ranked with BM25, the standard search formula: a passage scores higher when it contains the question's words, especially rare ones, without being padded with unrelated text. The top four are kept.

A score alone is not enough. The assistant also asks: how much of the question does the best passage cover? Each question word counts by how rare it is in your sources, so a word your sources never use (like “price” or “partnership”) weighs the most. If coverage is below the strictness you chose, the question goes to the moderator queue.

The answer is made of up to two sentences from the top passages that share the most words with the question, in their original order and wording, with the document and heading shown underneath. Nothing is paraphrased unless you ask for the optional AI wording.