Length changes the ranking problem
An operating manual may repeat the word inspection hundreds of times, while a short acceptance note contains the specific inspection decision the reader needs. Counting matches alone would favor the manual. The probabilistic relevance framework behind BM25 incorporates term saturation and document length considerations, addressing the fact that another occurrence of a word does not always add the same amount of evidence. This is useful background for evaluating search over mixed organizational document collections.
Choose the unit of retrieval
The proposed ABOLTUS approach should evaluate both document results and passage results. A long manual can be represented by meaningful sections while still preserving the parent document title and version. For a question about one equipment inspection, the result might identify a short acceptance note first and show a relevant manual section as supporting context. The interface should not hide the distinction between a whole document match and a passage extracted from a much larger source. Otherwise, users can mistake a strong local match for relevance throughout the document.
Measure relevance across formats
A practical test collection would mix short emails, large handbooks, spreadsheets and scanned attachments for the same topics. Reviewers could judge whether document length changes which useful evidence is found. BEIR supports the broader lesson that retrieval methods behave differently across datasets; its published findings do not establish a winning configuration for this particular organization. ABOLTUS would need local judgments before choosing length normalization settings. The objective is a reliable route to the relevant evidence, not a ranking system that rewards whichever department writes the longest files.