Find notes semantically similar to a given note using embeddings. Returns paths, similarity scores, and available blocks.

Build a multi-level connection graph starting from a note, showing how notes are semantically connected.

Semantic search over the vault. Returns an envelope, not a bare array: mode names the engine that answered ("semantic" is real, "keyword" means the embedding model failed to load and results will miss anything phrased differently), coverage says how many notes were actually searched out of the vault total, results holds the matches, and warning appears whenever the answer should not be read at face value. An empty results with mode "semantic" and full coverage means the vault really has nothing closer; an empty results with mode "keyword" or nonzero coverage.unsearchable means the tool was partly blind and you must not report it as an absence. Typical relevant matches score ~0.4-0.75; lower the threshold to widen recall.

Turn a wikilink into a real file path, the way Obsidian would. Accepts bare text ("Throughline"), full link syntax ("[[Throughline|the CRM tracker]]"), or a path, and honors aliases: frontmatter, so a tracker filed as 2026-06-23-throughline-workstream.md resolves from its codename. Use this whenever you read a [[link]] in a note and need the actual file, instead of guessing at the filename or falling back to semantic search, which finds notes that are SIMILAR rather than the one that was actually referenced. Returns null when nothing resolves, which means the note genuinely does not exist in this vault.

Which notes explicitly link TO this one. This answers a different question from semantic search: backlinks are edges the author wrote by hand, so they show what the vault has decided this note is load-bearing for, regardless of whether the prose is similar. Use it to judge how important a note is, to find every place a decision is cited before changing it, and to trace how a concept actually gets used. Set include_outbound to also get what this note links to.

Find wikilinks that point at notes which do not exist. Git proves two machines agree on what is COMMITTED; it is silent about a note written outside the vault folder or written on another machine and never committed, and both fail the same silent way, as a link that resolves to nothing. Results are ranked by how many DISTINCT notes reference each missing target, because one note pointing at an unwritten note is an ordinary forward reference while six pointing at the same target means the vault treats it as real and it is either a concept that never got a home note or a note this machine cannot see. Run at session start alongside check_search_health.

Positive control for retrieval. Asks the index for notes that are known to be there, by their own titles, and reports whether they come back. Use this at session start, before trusting any empty search result, and any time a vault has been quiet or moved between machines. Returns alive (false means every empty result from this server is untrustworthy), verdict (a plain-language line written to be read aloud), mode, coverage, and the individual probes. A retrieval tool that can return nothing must be able to prove it can still see.

Find nearest neighbors for a given embedding vector. Useful for custom similarity searches.

Retrieve the full content of a note, optionally with specific blocks/sections extracted.

Get statistics about the Smart Connections knowledge base (total notes, blocks, embedding model, etc.).