The tradeoff is hardware. Local models are slower and less capable than GPT-4 class APIs, so this is a privacy decision as much as a quality one. Below is what actually runs on 16GB and 32GB machines.

Why go local at all?

Three reasons show up again and again: client confidentiality, working offline on a plane, and predictable cost. A consultant handling contracts cannot paste them into a hosted endpoint without a data agreement. A local model sidesteps that entirely.

The catch: you are managing model files, embedding dimensions, and VRAM. It is not a one-click setup like pasting an OpenAI key.

What do you need installed?

1. Install Ollama for Windows, macOS, or Linux and confirm the service is running.
2. Pull a chat model: ollama pull llama3.1:8b for a 16GB machine, or qwen2.5:14b if you have 32GB.
3. Pull an embedding model: ollama pull nomic-embed-text, which the plugin uses for indexing.
4. In Obsidian Copilot settings, choose Ollama as provider and set the base URL to http://localhost:11434.
5. Select your chat model and set the embedding model to nomic-embed-text.

Run a test message in the sidebar. If it hangs, Ollama usually is not started, or the model tag is misspelled.

Which model should you pick?

The 8B tier is the sweet spot for vault Q&A. Smaller models miss nuance in long notes; larger ones make your fan spin without a huge accuracy jump for this use case.

How does local indexing differ from cloud?

In cloud mode, embeddings are generated by the provider and the vectors live in your vault folder. Locally, nomic-embed-text runs on your CPU or GPU and produces the same kind of vector store, still stored as a local file.

Re-index after switching embedding models. Vectors from different models are not compatible, and mixing them silently wrecks retrieval quality — you will get confident answers about the wrong notes.

What are the real limitations?

An 8B model will not summarize a 3,000-word research note as cleanly as a frontier API. It forgets context in long threads and sometimes invents a citation to a note that does not exist. Verify anything it claims you wrote.

Ollama also holds RAM between calls, so a 14B model plus Obsidian plus your browser on a 16GB laptop will swap and stutter. Match model size to free memory, not to the biggest number that downloads.

Who is the local setup for?

Anyone with confidential notes, or who works offline, and has at least 16GB of RAM. If you just want sharper answers and do not care about privacy, a hosted key still wins on quality. Local is the privacy-first path, not the best-answers path.

Quick Summary

• Obsidian Copilot works fully offline when paired with Ollama at localhost:11434.
• Pull a chat model plus nomic-embed-text for embeddings; re-index after any embedding model change.
• llama3.1:8b fits a 16GB machine; qwen2.5:14b needs around 10GB of VRAM or RAM.
• Local models are private but weaker than hosted APIs — verify every claim they make about your notes.

Setup takes about five minutes. The part worth understanding is the difference between chat mode, which answers from general knowledge, and vault mode, which reads your markdown files. Mixing them up is why people say the answers feel wrong.

What is Obsidian Copilot?

Obsidian Copilot is an open-source plugin maintained for the Obsidian note app. It adds three things: a chat sidebar, command-palette actions like summarize and rewrite, and a retrieval layer that indexes your notes for Q&A.

It is not a hosted product with its own model. It is a front end. You bring the API key and pay the model provider directly.

Is Obsidian Copilot free?

The plugin itself is free and open source. You still pay for model calls.

With a local model through Ollama, the whole stack costs nothing but your electricity and RAM. With a hosted key, light use runs a few dollars a month; heavy vault indexing of a 5,000-note collection can spike token costs, so check your provider dashboard the first week.

How do you install and connect a model?

1. Open Obsidian settings, go to Community Plugins, and turn off Restricted Mode.
2. Search for Copilot and install it, then enable it.
3. Open the plugin settings and pick a provider — OpenAI, Anthropic, Gemini, or Ollama for local.
4. Paste the API key and choose a model such as gpt-4o-mini or claude-3-5-sonnet.
5. Click Test to confirm the connection returns a response.

For Ollama, the base URL stays at http://localhost:11434 and you type the model tag you already pulled, for example llama3.1:8b. No key needed.

What is Vault QA and how do you turn it on?

Vault QA is the mode where Obsidian Copilot searches your notes and answers from them. Enable it, then kick off indexing from the plugin settings.

Indexing embeds each note and stores the vectors locally as a file in your vault folder. On a 2,000-note vault this takes a few minutes and produces a file you should add to your sync-ignore list so devices do not fight over it.

Chat mode vs vault mode

Does it send your notes to the cloud?

In vault mode, yes — the relevant note chunks go to whichever provider you configured. That is how retrieval works; the model cannot read files on your disk unless they are sent as context.

If your notes contain client data or anything sensitive, use a local Ollama model. Then nothing leaves the machine. Hosted keys mean the text you query travels to that vendor’s servers.

Who is this for, and who should skip it?

Writers and researchers with a big markdown vault get real value from asking “what did I decide about the pricing model” and getting an answer pulled from their own notes. It suits people comfortable managing an API key.

Skip it if you want a zero-config tool, or if your vault is under a hundred notes — the indexing overhead is not worth it, and plain search finds things faster.

Quick Summary

• Obsidian Copilot is a free, open-source plugin that adds AI chat and note Q&A inside Obsidian.
• The plugin is free; you pay for model calls or run it free with local Ollama.
• Vault QA indexes your markdown and answers from your notes; chat mode uses general knowledge.
• Vault mode sends relevant note chunks to your model provider — use a local model for sensitive data.