BM25

BM25 estimates how often query terms occur and how distinctive they are across the archive. A term frequent in one passage but rare across the collection can give that passage a stronger ranking signal.

Document length is part of the calculation. An accidental occurrence in a very long file is not treated like dense use in a short, focused passage; that is more balanced than raw word counting.

BM25 is useful for exact legal clauses, product codes, customer numbers, filenames, and technical terms. When the user knows the wording in the document, a direct lexical signal can be more informative than semantic proximity.

Synonyms and different languages are its natural limit. Leave and izin, or collection and tahsilat, can refer to the same idea without sharing a token.

KobiGPT’s retrieval layer can evaluate a sparse lexical signal alongside a dense embedding signal. That preserves names and codes while improving coverage for natural-language questions expressed in different ways.

A hybrid ranking is still not evidence by itself. Even a passage from the right file may miss the condition in the question; checking the source and its freshness remains part of safe answering.

A useful way to evaluate bm25 is to follow one real question from the source document to the final answer. Record which file was selected, what context reached the model, and what a reviewer would need to verify. This turns a definition into an operational check and makes the result comparable across teams.

The same check should include ownership and change management. Decide who updates the relevant documents, how an outdated result is reported, and which access boundary applies. KobiGPT can provide the assistant and the cited document context, but the organisation still owns the source material, permissions, and the decision made from the answer.