Mini-RAG v2.2.0 Release Notes
Features
-
Graph-augmented retrieval
After the normal vector search, Mini-RAG expands the context with excerpts from notes that are linked to or from the top results, following Obsidian[[wikilinks]]and backlinks. This improves multi-hop and associative questions, where the answer lives in a linked note rather than the one that best matches the query text. Notes are only expanded into if they're part of the current context (the note or folder the chat was opened on). -
Embedding cache (new setting, off by default)
Note embeddings can now be cached to disk so that re-opening a note or folder is near-instant instead of re-computing every time. The cache is sharded per folder under.obsidian/plugins/mini-rag/cache/, mirroring the vault structure. Each note's entry is keyed by a hash of its content, the chunking version, and the embedding model, so editing a note or switching models re-embeds only what actually changed. Entries for deleted notes are pruned automatically. -
Batched embeddings
Embedding requests are now sent to Ollama in batches of 32 rather than as a single large request. This prevents Ollama from running out of memory and crashing (previously seen as an "Embedding model not found" error) when indexing large notes or folders.
Full Changelog: 2.1.0...2.2.0