Everyone has a vault full of beautiful corpses. Notes you wrote at 2am, PDFs you highlighted, project logs, character sheets, half-finished essays. Then you install an AI plugin and it gives you glorified Ctrl-F with a chatbot face. That is not RAG. That is autocomplete wearing a trench coat.

The missing piece in 2026 is not a bigger model. It is local graph retrieval that understands relationships, runs entirely on your machine, and does not send your journal to someone else’s GPU farm.

Most plugins die because they rely on simple vector similarity. Ask “how does my note on burnout contradict my Q1 goals” and the bot returns three notes that share the word “tired”. That is mathematically correct and intellectually useless.

What works now is different. It is hybrid. Vector search for proximity, knowledge graph for structure, local reranking for precision, all inside Obsidian.

1. Why your current setup feels dumb

Standard RAG chops your notes into 500-token chunks, embeds them with whatever model was trendy, dumps them into ChromaDB. When you query, it finds the closest vectors. The problem is closeness is not meaning.

A researcher in the Obsidian forum described it perfectly in January. He asked his vault how a methodology in Paper A contradicted results in Project B. The vector plugin failed because the chunks were not textually similar, even though they were logically connected. He built Neural Composer to fix it, integrating LightRAG directly into Obsidian so answers come from relationships, not just keywords.

That is the shift. In 2026, local RAG is not retrieval augmented generation. It is retrieval augmented reasoning.

2. The stack that finally does not suck

You need four pieces and they all run locally. No API keys unless you want them.

First, the brain. Ollama. Not because it is perfect, but because it is boring and stable in 2026. Install it, pull two models. For embeddings, do not default to BGE-M3 on a laptop. It is 1024 dimensions, multilingual, beautiful, and it will timeout your CPU every time. The Neural Composer maintainer tells CPU users to switch to nomic-embed-text. It is lighter, faster, optimized for RAG, and stays under the 60 second worker limit even on modest hardware.

Second, the extractor. For building the graph you need a model that can actually see entities. A tiny 3B model will hallucinate relationships. Use Qwen2.5 14B q4 or Llama 3.2 11B if you have an M2/M3 Mac or an RTX 3060 or better. On Apple Silicon it runs surprisingly well. If you are GPU-poor, use Gemini 2.5 Flash for the initial ingest. The cost for 1,000 average notes is less than fifty cents, then you go fully local for queries.

Third, the store. Forget Pinecone. Use LanceDB or the built-in LightRAG store that lives in a .neural_memory folder inside your vault. That portability matters. When you sync via Git or iCloud, your graph travels with your notes, not in some hidden AppData folder.

Fourth, the plugin. In early 2026 there are two that actually work. Smart Connections still does local embeddings and semantic search, and it never phones home. Neural Composer adds the graph layer. It auto-starts a LightRAG server when Obsidian opens, kills it on close, supports PDFs and DOCX, and gives you hybrid search. It is free and open source.

That combination is the missing piece. Vector for “find me notes like this”, graph for “show me why these two ideas fight”.

3. How to build it without crying

Install Ollama. Open terminal.

ollama pull nomic-embed-text ollama pull qwen2.5:14b

Install Neural Composer via BRAT. Point it to your lightrag-server. On Windows, use the full path to the exe in your uv tools folder, not a venv shim. If you get “Server failed to respond in time”, it is almost always the embedding model timing out.

Right click your main folder, choose Ingest into Graph. The first run will take time. It is not embedding, it is extracting entities and relationships. You will see nodes like “Project Phoenix”, “burnout”, “methodology”, and edges labeled “contradicts”, “depends on”, “causes”.

Do not let the plugin regenerate your.env every time. Use the built-in editor under Settings, Neural Composer, Review.env and Restart. Add your timeouts there. Manual edits get overwritten because the plugin manages the file authoritatively.

The smell while this runs is warm electronics and that faint sandalwood from the candle I keep by my monitor. My laptop fan sounds like a cat purring. Let it finish.

4. Chunking is where most people lose

Do not use fixed 512-token chunks. Obsidian notes have structure. Use heading-aware chunking. A chunk should be one idea, usually one H2 or H3 section plus its list items. Neural Composer does this automatically. Smart Connections lets you set it.

For PDFs, extract text with layout preservation. LightRAG handles PDFs natively now. If you have tables, use the multimodal embedding path. The 2026 embedding comparison shows no single model wins everywhere, but for local use, nomic-embed-text for text and bge-m3 for hybrid text plus images is the practical balance.

Keep your chunk overlap at 15 percent, not 50. Overlap creates duplicate vectors and confuses the reranker.

5. The query pattern that feels like magic

Once ingested, stop asking keyword questions. Ask relationship questions.

Instead of “notes about sleep”, ask “what in my vault explains why my sleep notes contradict my productivity system”. The graph traverses from “sleep hygiene” node to “late night writing” to “Q1 goals” via an edge created during ingest.

The plugin does hybrid retrieval. It pulls precise file snippets for citations, then pulls global graph context for synthesis. That is why it can answer synthesis questions that break pure vector search.

Enable local reranking in v1.1.x. It uses a tiny cross-encoder that runs on CPU and reorders the top 20 hits. The difference is night and day. Without reranking you get recall. With it you get relevance.

6. Privacy and cost, because you care about both

Everything lives in .neural_memory in your vault. No cloud. The January 2026 guide to local RAG with Ollama emphasizes this exact architecture for enterprise environments: data privacy, zero API costs, full infrastructure control.

If you choose to use Gemini 2.5 Flash for the heavy ingest, you are sending text once, not continuously. After that, queries are local. Ongoing cost is pennies per month or zero.

On a MacBook Air M2 with 16GB RAM, expect 8 to 12 tokens per second for qwen2.5:14b. That is fast enough for chat. On a Windows machine with RTX 4070, it flies. On CPU-only, stay with nomic-embed-text and use a smaller 7B model for queries, accept that graph building will be overnight work.

7. What breaks and how to fix it

Timeouts. If embedding times out, lower EMBEDDING_BATCH_NUM to 1 and raise EMBEDDING_TIMEOUT to 600 in the plugin env editor. Do not edit the.env file directly.

Messy graphs. If your graph is full of generic entities like “thing” or “idea”, your extractor model is too small. Do not go below 7B parameters for graph extraction or the graph will be messy.

Server not starting on Windows 11. Point to the actual lightrag-server.exe in AppData Roaming uv tools, not the shim. Then restart Obsidian as administrator once to allow the port binding. (you should try linux, by the way.)

Irrelevant answers. Turn on hybrid mode, not pure vector. Pure vector is fast and dumb. Hybrid is slower and smart.

8. The workflow I actually use at night

I open Obsidian at 11pm. Neural Composer heartbeat turns green in the status bar. I type into the chat: “summarize the tension between my recent observations on local patterns and the research on personal ambiguity”. It returns three notes with direct quotes, then a synthesis paragraph that references a PDF I imported last month. Each citation is clickable back to the exact heading.

I curate the graph once a week. The 2D Sigma.js view shows duplicate entities. I merge them. I add manual edges with the Relationship Weaver. That curation is the secret. AI builds the first draft of your second brain. You edit it.

The whole vault is 4,200 notes. Initial ingest with Gemini 2.5 Flash cost 43 cents. Daily queries cost nothing. The data never left my laptop.

That is what was missing for two years. Not another cloud assistant. Not another subscription. A local system that treats your notes like a living network, not a bag of words.

You have the vault. You have the ideas. Now you have the piece that makes them talk to each other without asking permission from the internet.

Install it tonight. Ingest a small folder first. Ask it something cruel and specific. Watch it find the connection you forgot you wrote. Enjoy the newfound clarity and your new on-demand epiphany mainline.

If you want the step-by-step deep dive, check it out below:

From Notes to Knowledge Graph: Local RAG for Obsidian

Check out the other latest drops as well:

The Stingray Survival Kit: Hardware + Software Methods to Detect Fake Towers

Be Anywhere But Here - Fake Your iPhone Location Convincingly