Wiki Retrieve is the search skill of the claude-obsidian plugin. It splits wiki pages into chunks, adds a short page-level prefix to each, builds a BM25 index and returns paths and snippets for a query.
Everything it derives is stored under the vault’s .vault-meta folder and your notes are never changed. Reranking with a local embedding model is optional, and sending page text to a remote service needs your explicit consent.
What You Get
- A local contextual BM25 index of the wiki, kept as disposable cache files under .vault-meta
- Query results with page paths, snippets and an optional explanation of the ranking
- Optional rerank with a multilingual or a smaller English Nomic embedding model
- Diagnostic commands for index stats, direct queries and rerank previews
- A stable error and rebuild command when the index is missing or corrupt
Common Use Cases
- Finding the passages in a large vault that match a question
- Building the search index that the wiki-query skill uses
- Comparing lexical search with reranked results on your own notes
- Measuring recall and ranking quality on a set of labeled questions
What Makes It Unique
The index is local, deterministic and built with the Python standard library, so a basic setup needs no model download. Reranking falls back to the original BM25 order as a whole when any embedding step fails, so score scales are never mixed. The skill also checks that every path stays inside the vault, rejects symlink escapes and stale hashes, and requires consent before page text leaves the machine.
Under The Hood
The skill loads its SKILL.md and runs shell commands that call four Python scripts from the installed plugin: one to split pages and write prefixes, one to build the BM25 index, one to query it and one to rerank. They read the wiki folder and write chunk and index cache files under the vault’s .vault-meta folder. Reranking calls a local Ollama service on localhost over HTTP. An option exists to send page bodies to the Anthropic API, and another to use a remote Ollama endpoint; both need your consent and an explicit flag. It does not fetch web pages.