QMD
An on-device search engine for notes, transcripts and docs whose query mode can send each candidate document through TypeSafe's System One endpoint for an independent relevance score before ranking.
An on-device search engine for notes, transcripts and docs whose query mode can send each candidate document through TypeSafe's System One endpoint for an independent relevance score before ranking.
With the rerank model set to typesafe:<model>, every candidate chunk becomes its own request: the state is the search query, the document text and the document title when one exists, and there is one noul question keyed relevant that asks whether the document contains information answering the query. The returned probability is that chunk's rerank score, and the fused BM25 and vector candidates are sorted by it. At most two requests are in flight and one call carries at most 64 documents.
Does the document contain information that answers or directly addresses the search query?
The state is the search query plus one candidate document and, when available, its title. Each candidate is asked in its own request, so the excerpt below stands in for one of them.
qmd expands the query and runs BM25 full-text search and vector search over the indexed notes and documents.
Reciprocal Rank Fusion merges the two ranked lists into one candidate set for the query.
When the rerank model is a typesafe: model, each candidate is sent as its own request with the relevance noul question and the configured model name.
Chunks are sorted by the returned probability. A missing credential, invalid response or timeout fails the query instead of silently keeping the fused order.
A local-first search engine borrows a hosted decision model for one narrow step. Independent per-document questions make the scores comparable and let one document's text stop influencing the next document's answer, so a large candidate set can be scored in parallel without turning the ranking into a single shared prompt.