Skip to main content
search asks a mount’s own search to rank its records for a query: Dify’s retrieval API, Chroma’s and Qdrant’s vector queries, mem0’s memory search and LanceDB’s vector search. Use it when meaning matters; grep and rg still match exact text.

Setup

search has no account config. It searches the mounts it is pointed at, so register it on a workspace that mounts a store:
Mounting a store does not register search. For YAML registration, see the CLI overview:

Supported commands

Run search --help for the supported flags and operands. This list is generated from Mirage’s registered command tree.

Usage

Each hit prints its path, then its score where the store ranks, then the record’s text:
A file the session cannot see is never printed.

Mounts

Answer from a custom VFS

A custom VFS joins the list by defining search. It gets the path the line names and a SearchQuery: the text in query, and in options the top_k, threshold and method the line gives. It returns one (path, record) pair per hit, the hit’s workspace path and the text search prints; [] means no match and None declines. A VFS that ranks several paths together defines search_many(paths, query) as well. Pass the path on to the service, so a search under one folder ranks only that folder’s notes and no other folder’s hits use up top_k.
A search never passes the file cache. To keep results, the accessor holds a ResponseCache and returns it from caches(); the mount then expires it with its read policy and drops it after a write from outside Mirage.
await self.searches.get(key, fetch) answers from the cache or calls fetch.

Errors