org-semantic brings embedding search to org-mode. A vector embedding model reads each passage of your notes (what is internally called a chunk) and turns it into a point in a high-dimensional vector space. Passages saying the same thing sit close together, whatever words they used and even language (English, French, etc.) if you used a multi-language embedding model. Your query becomes a vector in the same space, and the score simply reflects the projection of the query vector onto the indexed points. So an org note in your collection can match a question with which it has no word in common; in fact, an English question can have a best match to an Italian note. Traditional search is there too. The words themselves go into a second index, ranked by BM25 with per-language stemming. When you run a search, you can choose and switch which of the two rankings (semantic vs. lexical) to use. The search itself is a single Rust binary from the same repository. This package is its client. It starts the binary, keeps it resident over a JSON-RPC pipe, and presents the results in a buffer, where each candidate match opens the note at its line. The embedding models run entirely on your machine, and no information from your notes or your searches is ever transmitted anywhere. Each model is downloaded once and cached. To start: - Set `org-semantic-vault-root' to the directory of your notes. - M-x org-semantic-find searches them. When the binary, the model or the index is missing, it says so and offers to get it. - M-x org-semantic-reindex updates the index. `org-semantic-auto-reindex-mode' updates it when you save a note. - M-x org-semantic-doctor says what a vault has and what is wrong. The manual is at https://alberti42.github.io/org-semantic/.