Smart Second Brain for Obsidian(Free & Offline)

A privacy-first “second brain” for Obsidian can run entirely offline by pairing the Obsidian plugin Smart Second Brain with locally hosted large language models (LLMs). The setup turns a personal Obsidian vault into a searchable knowledge base: users can chat with their notes, ask questions, generate summaries, and get suggestions with links back to the exact notes that informed the answers—without sending content to the cloud.

The workflow starts by installing an offline LLM runner called OMA (from oma.com). OMA provides the ability to run LLMs locally on Mac, Linux, and Windows, including models such as Llama2, Gemma, and Dolphin Mixtral (as listed in the transcript). After OMA is installed and running, the next step is adding the Obsidian plugin Smart Second Brain. Once enabled, the plugin can auto-start with Obsidian and lets users exclude specific files or folders from indexing to avoid long waits—important because indexing time scales with vault size.

Smart Second Brain then connects to OMA for two key model roles. First is the embedding model, which indexes the vault so the system can retrieve relevant notes when a question is asked. The transcript highlights “mxb y embed large” as a recommended embedding choice. Second is the chat (language) model, which generates the final response. The plugin also includes retrieval controls such as similarity and “creativity,” plus an option to use third-party services (notably OpenAI), though the demonstration focuses on local OMA for offline privacy.

After installing the chosen models, Smart Second Brain initializes and indexes the vault. In the demo, a vault with roughly 250 nodes indexed in minutes, but larger vaults can take longer. Once indexing completes, the “octopus” toggle determines whether answers draw from the vault’s indexed notes. With the toggle off, the assistant replies using its general training. With the toggle on, it retrieves relevant notes and can create clickable references back into the vault.

The transcript also surfaces limitations and tuning behavior. When retrieval fails, the assistant may produce answers that don’t match the vault or may generate links to notes that don’t exist—an issue the creator hopes to see fixed. Users can adjust similarity and creativity to improve retrieval quality. Switching chat models (e.g., from GMA to Mixtral) forces reindexing and changes answer accuracy; the demo reports better alignment after reindexing with Mixtral.

Model choice affects both quality and hardware demands. The transcript compares Mixtral’s performance to GPT-3.5-class results, but notes it is resource-heavy (around 25GB), requiring a powerful PC. For the best balance, the demonstrator settles on Llama2 as the chat model paired with “mxb y embed large” as the embedding model, citing improved answers and correct references.

Beyond Q&A, the plugin supports saving chats back into the vault as notes, and deleting chats when needed. Overall, the core value is a conversational Obsidian experience grounded in personal notes—fast enough to iterate, configurable enough to tune, and private enough to keep everything local.