From your question
to the quoted line

Every answer this engine gives is a line that already existed in one of your documents. It has no language model inside it to compose a sentence, so it cannot compose a wrong one. It can find, rank, gate and quote — and each step is plain enough to read.

The path from a question to a quoted line has five steps. Each is deterministic code — the same question against the same records returns the same line every time — and each is small enough to read. The whole engine is 5,639 lines across 23 modules actual, and none of them calls a language model.

A question that names a specific job has to be answered from the records that mention that job. In a real filing cabinet that is harder than it sounds: rate sheets are full of the common words in your question — cost, metre, bore — and they crowd the one document with the rare name out of a short list. So the engine searches the whole question first, then each meaningful word on its own, four times deeper. The passage carrying the rare name stays in the pool.

When the pool is ordered, each word in your question is weighted by how rare it is among the candidates: matching the job name is worth far more than matching "per". That weighting is estimated at question time; nothing is fitted in advance.

Three methods score every passage in the pool. BM25 and TF-IDF are lexical: they reward passages that contain the words you used, weighted by how unusual those words are. The third is an embedding — passage and question are each turned into a vector by a small model, and the closeness of those vectors is the score. That model, BAAI/bge-small-en-v1.5, runs on your machine through fastembed — the only model in the engine, and it never composes text.

The three scores sit on different scales, so each is min-max normalised and combined by fixed weights — configuration, not learned — into one fused score. The engine also records, for every passage, what fraction of the question's meaningful words it contains: that is the coverage the gate uses later.

The reranker is a short piece of arithmetic, not a model. It starts from the fused score and adds a fixed credit for coverage, a bonus if the passage contains your question as an exact phrase, and a bonus for each literal number from your question that appears in it. The candidates are then sorted by rarity-weighted coverage, rerank score breaking ties — the winner is the passage that covers your question's distinctive words best.

Before anything is quoted, the winner goes through the gate. It must clear a score floor and a coverage floor, both. If it fails either, the engine says so and quotes nothing. That decision has its own page: it is the most important one the engine makes, and the one time it was measured it was not good enough.

The winning passage is split into its lines. Each is scored against the question — content-word overlap plus a bonus for matching numbers — and the top three with any relevance are kept, in document order, and returned exactly as written. Nothing is paraphrased, summarised or completed.

Each citation carries the document id, the source path and a locator: line numbers for text, a page number for a PDF, a cell reference for a spreadsheet. A second citation is added only if the second-best passage clears the gate on its own; a low-confidence chunk is never cited to pad the answer. Three lines and two citations are the ceiling, and the ceiling is the promise: the answer is a line out of your document, not a sentence about it.