Obsidian LLM Wiki Local: Karpathy's Concept Graphs Go 100% Private

Zero-Network Pipeline Design

The architecture centers on a local-only inference loop that never exposes note content to external APIs. A Python watcher monitors your Obsidian vault, triggering Ollama-hosted LLMs to analyze Markdown semantics.

Concept Extraction Flow

1. Chunking: Parses Markdown into semantic blocks (paragraphs, lists) while preserving frontmatter
2. LLM Analysis: Prompts local model to identify concepts (abstract ideas) vs. keywords (literal text)
3. Link Resolution: Matches extracted concepts against existing note titles and previous extractions
4. Vault Mutation: Writes bi-directional links directly into source Markdown, compatible with Obsidian's graph view

Privacy Architecture: Unlike vector-RAG systems that embed content into retrievable vectors, this approach writes connections as plain text links, making the "AI layer" removable without data loss.

The "RAG Alternative" Philosophy

Traditional RAG is reactive: you query, it retrieves. This system is proactive: it writes connections into your notes as you create them, effectively turning your LLM into a co-author rather than a search engine.

• Concept-Centric Linking: Uses LLM reasoning to connect "Transformers" (AI) with "Attention Is All You Need" (paper) even when literal keywords don't overlap
• Organic Growth: The wiki structure emerges from content semantics rather than manual curation or rigid folder hierarchies
• Git-Native Versioning: Treats knowledge evolution as code—every AI-suggested link is a diff you can review, revert, or merge

Local-First Constraints as Features

By mandating Ollama, the system forces optimization for quantized models (4-bit/8-bit), resulting in surprisingly efficient concept extraction that runs on consumer hardware (M1 Macs, RTX 3060s). The constraint eliminates the "API anxiety" of sending personal notes to OpenAI/Anthropic.

Inference Latency by Model Tier

Benchmarked on M2 Pro (32GB). Performance scales linearly with vault size but parallelizes across files.

Scalability Ceiling

The current architecture hits practical limits around 10,000 notes with 7B models, primarily due to context window constraints during cross-note linking. For larger corpora, the tool supports sharded processing—processing recent notes daily and running full-vault link reconciliation weekly.

Limitations

• Cold Start: Initial processing of a 1,000-note vault takes 20-40 minutes depending on model size
• Hallucinated Links: Smaller models (3B) occasionally suggest spurious connections requiring manual cleanup
• English-Centric: Concept extraction quality degrades significantly for non-Latin scripts with base Ollama models

Obsidian Integration

Works natively with Obsidian's core plugins—suggested links appear as standard [[WikiLinks]], compatible with Graph View, Backlinks panel, and Dataview queries. No proprietary JSON formats or lock-in.

Ollama Model Compatibility

Deployment Patterns

• Desktop: Python script + Ollama desktop app (macOS/Windows/Linux)
• Homelab: Dockerized Ollama + cron-triggered wiki updates
• Sync-Safe: Git-based conflict resolution when using Obsidian Sync or GitHub

Commercial Considerations

MIT licensed with no commercial restrictions. The 27 forks suggest active experimentation, including community adaptations for Logseq and Emacs Org-mode. No SaaS upsell—truly local-first.