9.6.2
9.6 Upgrade to v4 Transformers.js for better GPU support
- 9.6.1-2: Fix required Obsidian version.
We now better support local GPU for embeddings via Transformers.js. On my development system I achieve 50--60% GPU utilization. This also reduces the plugin size, shrinking main.js by over 12%.
- Migrated local tokenization and embedding engine to
@huggingface/transformers(v4.2.0), enabling dynamic WebGPU acceleration and clean fallback capabilities.- Changed quantization setting for default embedding model (Xenova/multilingual-e5-small) which now consumes 60% of the developer's GPU during re-embedding, confirming the WebGPU is working.
- Added a "Quantize local model" toggle setting for local (Transformers.js) embedding models, enabling power users to disable 8-bit quantization and run unquantized (fp32) models fully on the GPU (WebGPU) instead of falling back to CPU/WASM.
9.5Support Voyage AI and Gemini 2 embedding models
- 9.5.8: Updated upstream dependencies (#565), updated README, and completed missing entries in CHANGELOG.
- 9.5.1-7: Address some of the warnings from the Obsidian community automated checks.
- 9.5: Added support for
gemini-embedding-2models. Developers see here for changes to support migration. - See the full CHANGELOG.
This release includes support for Voyage AI as an embedding provider. This is a key step towards supporting Claude and others who do not bundle an embedding model with their main reasoning models. We now support:
- Local embeddings, via Transformers.js and Ollama.
- Cloud embeddings, via Gemini, Voyage AI, and Ollama.
This release includes the major features from the release 9 branch:
9.3 — Ollama custom headers
This release focuses on enhancing connectivity for local AI environments and modernising our core development foundations. We've introduced custom headers for Ollama, giving users greater control over their private infrastructure, while upgrading to TypeScript 6.0 and ESLint 10 to ensure the plugin remains robust and future-proof.
Custom Headers for Ollama
You can now configure custom HTTP headers for Ollama requests, and these headers can be securely held in your system's keystore.
- Flexibility: Connect to remote Ollama instances that require
Authorization(see Ollama documentation on API Keys) or custom headers. - Privacy: Maintain secure connections to your private infrastructure without sacrificing ease of use.
9.2 — Vault Hygiene and Stability
This release focuses on vault organisation and system stability. As your knowledge base grows, it naturally accumulates duplicate ideas and abandoned notes. We have expanded the Gardener to detect and resolve these issues automatically. Alongside these new capabilities, we have fixed memory leaks and implemented important security updates to the Model Context Protocol (MCP) clients.
Note: Upgrading from 9.1.x or earlier will trigger a full re-indexing of your vault on the next workspace load, which may consume AI tokens if you are using a cloud-based embedding model. Upgrading from 9.2.0 will not trigger index.
Gardener Semantic Merging
You can now use the Gardener to intelligently detect identical and duplicate topics in your ontology. Rather than just finding exact text matches, the Gardener uses a triple-detection system combining lexical, structural, and semantic vector checks to identify conceptual duplicates.
Crucially, this is designed with safety in mind. Merging topics uses Obsidian's intelligent parsing to rewrite links throughout your vault without breaking references. All discovered merges are presented in an interactive card, giving you full oversight before any files are modified.
Gardener Orphan Pruning
Knowledge bases naturally accumulate cruft over time. The Gardener now mathematically detects abandoned topic notes that have no incoming links (either via wikilinks or frontmatter). It will propose these orphans for archive or deletion, helping to keep your vault cleanly structured and focused on active concepts.
