LLM Automation For Obsidian Plugin

The plugin's surface is 40+ Obsidian commands, ribbon icons, side-pane views, and modals across five workflows: text & image generation, summarization & transcripts (article, single-video, and playlist/batch with a processing queue), learning (quizzes, flashcards with a database-backed spaced-repetition scheduler, and C# LINQPad coding exercises with local compile-and-run), knowledge retrieval, and path/tag management.

The provider layer is the defining design choice: five OpenAI-compatible providers behind one factory + service, so every LLM-calling feature is provider-agnostic. The proxy provider is explicitly designed to point at the openai-nim-proxy gateway, making this plugin both a standalone tool and a first-class client of that gateway.

The retrieval subsystem is a local hybrid RAG engine: FTS5 lexical search plus optional semantic search via local Ollama or remote Chutes embeddings, reciprocal-rank fusion, a token-budgeted evidence-pack builder, and a bounded multi-step KnowledgeAgent that can request more reads before answering with citations. A vendored FTS5-enabled sql.js wasm solves an environment-specific problem (Obsidian's bundled wasm lacks FTS5).

Key features

• Five OpenAI-compatible providers (OpenRouter, Chutes, Z.AI, Ollama, proxy) behind one factory.
• Text & image generation with per-run temperature/max-tokens/top-p overrides.
• Article, single-video, and playlist/batch summarization with a processing queue.
• Database-backed spaced-repetition flashcards (Mnemosyne-like scheduler).
• C# LINQPad coding-exercise generation with local compile-and-run (gated).
• Local hybrid RAG: FTS5 + optional semantic embeddings + RRF + bounded KnowledgeAgent with citations.
• Vendored FTS5-enabled sql.js wasm (Obsidian's bundled wasm lacks FTS5).
• 472 Vitest tests across 40 files with a full Obsidian API stub.

Stack

Gallery