repository·main·Indexed 25 days ago
https://github.com/glowingjade/obsidian-smart-composerAn Obsidian plugin that enhances AI-assisted writing by integrating vault content into AI context. It features @file mentions, semantic vault search (RAG), one-click document edits, and support for the Model Context Protocol (MCP). Compatible with various API providers including OpenAI, Anthropic, Gemini, and local LLMs via Ollama or LM Studio.
Smart Composer includes a Retrieval-Augmented Generation (RAG) feature that uses semantic search to find relevant notes from your vault to enhance AI responses.
To run a Vault Search answer, use the keyboard shortcut:
Cmd+Shift+EnterStandardize repetitive tasks by creating and reusing templates. In the chat view, type / to access your templates. You can also create templates from any selected text with one click.
You can define your own system prompts in the settings, which will be applied to every chat conversation to guide the AI's behavior.
To use Smart Composer in Obsidian, follow these steps:
Install via Community Plugins:
Community plugins and click Browse.Smart Composer and click Install.Configure AI Providers:
You can choose between two connection methods in Settings > Smart Composer:
Settings > Smart Composer > Providers for providers like OpenAI, Anthropic, or Gemini.Troubleshooting Installation:
If the plugin fails to load, ensure your Obsidian installer is up to date. If Settings > General > Check for updates doesn't resolve it, manually download and run the latest installer from obsidian.md.
The Contextual AI Assistant allows you to precisely control the information the AI uses for its responses.
Use the @ symbol in the chat to select specific context:
@<fname> to choose specific files or folders as your conversation context.You can expand the AI's context by adding non-textual information:
You can customize which AI models power your experience via Settings > Smart Composer > Providers.
For complete privacy and offline usage, you can run open-source LLMs and embedding models locally using Ollama, [LM Studio], or [MorphLLM].
The project uses Drizzle ORM with a PostgreSQL dialect. The configuration is defined using defineConfig from drizzle-kit. To maintain the database schema, ensure the schema path points to ./src/database/schema.ts and the dialect is set to 'postgresql'.
import { defineConfig } from 'drizzle-kit'
export default defineConfig({
dialect: 'postgresql',
schema: './src/database/schema.ts',
})smartComposerSettingsSchema defines the structure and validation for the plugin's configuration. It includes settings for LLM providers, chat models, embedding models, RAG (Retrieval-Augmented Generation) options, MCP (Model Context Protocol) server configurations, and chat behavior. If values are invalid, the schema uses .catch() to revert to safe defaults.The plugin provides several ways to interact with the Chat interface:
openChatView(openNewChat?: boolean): Opens the chat view. If openNewChat is true, it starts a fresh session. If an active editor is present, it attempts to include the selected block data in the context.activateChatView(chatProps?: ChatProps, openNewChat?: boolean): Low-level method to ensure the chat leaf is active and set to the correct state.addSelectionToChat(editor: Editor, view: MarkdownView): Specifically targets the current selection in the editor and adds it to the existing chat session, then focuses the message input.Use formatMessages() to prepare an array of RequestMessage objects for LLM API consumption. This function ensures compatibility with providers that require strict alternation between user and assistant roles by merging consecutive messages of the same role.
Key behaviors:
system role are collected, concatenated with double newlines, and placed at the very beginning of the message array.user messages in a row) are merged into a single message.string and ContentPart[] formats. If all merged parts are text-based, the resulting content is returned as a single concatenated string. Otherwise, it returns a ContentPart[] array.DatabaseManager handles the underlying storage. Use getDbManager() to retrieve the instance. This is an asynchronous operation that initializes the database (using PGLite). If initialization fails due to a PGLiteAbortedException, the plugin will prompt the user to update the installer.