Local AI for Obsidian: chat with your notes privately

Run Ollama on your computer, install an Obsidian community plugin that supports Ollama, and point it at http://localhost:11434. Your notes then get searched, summarized and questioned by a model on your own machine. Nothing is uploaded.

That matters more for notes than for most things. A vault is often years of journals, work notes and half-formed ideas. It's exactly what you don't want to send to a cloud service.

What you can do with it

• Ask questions across your vault. "What did I decide about the kitchen quote?" pulls the relevant notes and answers from them.
• Find related notes. Embeddings link notes by meaning, not just shared words, so you rediscover things you'd forgotten writing.
• Summarize and rewrite. Turn a messy meeting note into a clean summary, or a long note into bullet points.
• Draft from your own material. Outline a post using only what's in your notes.

Step 1: Ollama and two models

Install Ollama, then pull a chat model and an embedding model. The embedding model turns notes into vectors for search; it's small and fast.

Step 2: Pick a plugin

Obsidian's community plugins include several that work with local models. They fall into a few kinds:

• Chat plugins that open a side panel and talk to the model, often with the current note or selected notes as context.
• Semantic search plugins that embed your vault and show related notes or answer questions across it.
• Text tools that run a prompt on your selection, like "summarize" or "fix grammar".

Open Settings, then Community plugins, and search for "Ollama" or "local LLM". Check that a plugin lists Ollama or local models as a supported provider, that it has been updated recently, and read what it does with your data. Most let you choose a provider in their settings; pick Ollama and enter the address above.

If a plugin can't connect even though Ollama is running, the usual cause is Ollama refusing requests from the Obsidian app. Ollama has an OLLAMA_ORIGINS setting for this; the plugin's own docs normally say exactly what to set.

Step 3: Let it index your vault

Semantic search plugins embed every note once, then update notes as they change. On a GPU this is quick even for thousands of notes. On a laptop without one, expect the first pass to take a while, and leave it running.

Which model to use

Notes questions are mostly retrieval and summarizing, so you don't need a big model.

A lot of Obsidian users are on Macs. Here's a 16 GB base M4 from our calculator:

At around 13 tok/s for Qwen3 8B, that's comfortable for reading answers as they appear. Qwen3 4B at about 26 tok/s feels snappier for quick rewrites.

Make the answers good

1. Write notes the search can find. Clear titles and headings help retrieval more than any setting.
2. Ask it to cite notes. "Answer only from my notes, and name the note you used" cuts invented answers.
3. Raise the context length if the plugin sends several notes at once. See how much VRAM you need.
4. Turn off thinking for quick tasks. With Qwen3, add /no_think to a rewrite prompt and it answers straight away.

Without a plugin

Your vault is a folder of plain Markdown files, which makes it easy to use with other local tools. You can upload it to a Knowledge collection in Open WebUI, or run a script over every note, like tagging or summarizing a whole folder; see batch jobs with a local model. For how search-then-answer works under the hood, read chat with your documents.

Privacy check

Plugins can call any service they're built for, so check each one's settings point only at your local address. Then switch off Wi-Fi and ask a question. If it still answers, nothing is leaving your machine. See what stays on your machine.