LLM-Wiki

A local, LLM-maintained personal knowledge base. Drop documents in, watch an LLM compile them into a living, interlinked Obsidian wiki you can search and query.

Feel free to fork and don't forget to give it a Star ⭐️ for better reach!

Hello, I'm Nihar Shrotri, working as an AI Consultant. I'm currently pursuing my PhD in Artificial Intelligence and Machine Learning

Let's connect on LinkedIn for a Chat: https://www.linkedin.com/in/niharshrotri/

Built on the pattern Andrej Karpathy described in his LLM Wiki gist: instead of retrieving from raw documents at query time (classic RAG), an LLM incrementally compiles your sources into a structured, cross-linked markdown wiki that sits between you and the raw documents. The wiki is a persistent, compounding artifact — the cross-references are already there, the contradictions have already been flagged, the synthesis already reflects everything you've read.

You never write the wiki yourself. The LLM does all the grunt work: summarizing, cross-referencing, filing, bookkeeping. You bring the sources and ask the questions.

Runs 100% locally on Apple Silicon or anywhere Ollama works. No API keys, no cloud, no data leaving your machine.

What it does

# Drop files in (PDFs, markdown, HTML, DOCX, text)
wiki add ~/Documents/papers --recursive

# Watch Qwen3 read them and build an interlinked wiki
wiki ingest

# Ask questions — it searches the compiled wiki and cites its sources
wiki query "what's the main argument about X?"

# Health-check the knowledge base
wiki lint --fix

# Browse the whole thing in Obsidian (graph view, backlinks, everything)
open wiki/

Every ingest produces a cluster of sources/, entities/, and concepts/ pages with YAML frontmatter and [[wikilinks]] between them. Every query pulls the top-ranked pages via hybrid BM25 + vector + LLM-rerank search, then synthesizes a cited answer. Every lint run catches broken links, orphan pages, malformed frontmatter, and (optionally, using the LLM) contradictions between pages.

Features

Core capabilities

• Incremental ingest — drop a file, run wiki ingest, get 8–15 cross-linked wiki pages
• Structured extraction — Qwen3 identifies entities (people, orgs, models), concepts, and key takeaways per source
• Smart merging — re-ingesting related sources updates existing entity/concept pages instead of overwriting them, preserving provenance
• Hybrid search — BM25 full-text + vector embeddings + LLM reranking (all local, via QMD)
• 3-way query scope — Wiki (thematic answers from LLM-compiled pages), Raw (exact lookups in original documents), or Hybrid (both)
• Intent classification — casual messages ("hi", "thanks") skip retrieval and get a quick reply, saving ~30 seconds per chitchat turn
• Cited synthesis — queries return markdown answers with [[wikilinks]] pointing to the pages that support each claim
• Write-back — save good answers as new synthesis/ pages with --save-as, so your explorations compound in the knowledge base
• Wiki linting — automated health checks for broken links, orphans, malformed frontmatter, noise in sources, and (with --deep) LLM-powered contradiction detection between pages
• Auto-fix — most stylistic issues resolve with one command
• Auto-reindex — search index refreshes automatically after ingest and lint; new pages are queryable immediately

Web UI

A full web interface at http://127.0.0.1:8000 after wiki serve:

• Dashboard — project stats and recent activity
• Sources — list, inspect, delete, or re-ingest sources with one click
• Ingest — drag-and-drop upload, live progress log, persistent jobs that survive tab close and server restart
• Jobs — history of all ingest runs with live progress bars and error details
• Query — chat-style interface with streaming synthesis, scope toggle, save-as-synthesis button
• Lint — interactive lint report with one-click auto-fix
• Graph — D3 force-directed visualization of the full wiki, color-coded by page type

Supported input formats

.pdf · .md · .html · .docx · .txt

Obsidian integration

The wiki/ folder is a ready-made Obsidian vault with:

• Color-coded graph view (sources, entities, concepts, synthesis each get their own color)
• YAML frontmatter compatible with the Dataview plugin
• All cross-references as native [[wikilinks]] so backlinks, outgoing-links, and graph traversal all work

Architecture

Three layers, per Karpathy:

┌───────────────┐   ┌───────────────┐   ┌───────────────┐
│   raw/        │ → │   LLM Agent   │ → │   wiki/       │
│ Your docs     │   │  (Qwen3-14B)  │   │ Markdown,     │
│ (immutable)   │   │               │   │ auto-linked   │
└───────────────┘   └───────────────┘   └───────────────┘
                          │                     │
                          ▼                     ▼
                    ┌───────────────┐   ┌───────────────┐
                    │ schema/       │   │   Obsidian    │
                    │ AGENTS.md     │   │  graph view   │
                    │ (the rules)   │   │  + editing    │
                    └───────────────┘   └───────────────┘

• raw/ — your source documents. Immutable. The agent reads but never modifies.
• wiki/ — LLM-maintained markdown. One folder per page type (sources/, entities/, concepts/, synthesis/) plus auto-generated index.md and log.md. Open this in Obsidian.
• schema/AGENTS.md — the conventions file. Tells the LLM how to format pages, when to merge vs create, how to cite, how to handle contradictions. Edit as your preferences evolve.
• .wiki/ — internal state: SQLite ingest history, QMD search index, config. Git-ignored.

The ingest pipeline

Each source goes through three LLM passes:

1. Extraction (thinking mode on) — Qwen3 reads the source and returns structured JSON: summary, key takeaways, named entities, concepts, tags.
2. Page drafting (streaming, thinking mode off) — one call per entity/concept. Draft a new page from scratch, or merge new information into an existing page (preserving prior content, updating dates, appending to sources: frontmatter).
3. Source summary — write the sources/<slug>.md page listing every wiki page touched by this source for provenance.

After the three passes: index.md is rebuilt, log.md is appended, and QMD's search index is updated automatically.

The query pipeline

1. Hybrid search via QMD — BM25 full-text + vector similarity + LLM reranker, all local
2. Top-K page hydration — load full content of the top 5–8 hits
3. Synthesis — Qwen3 writes a cited markdown answer using [[wikilinks]] to reference the pages
4. (Optional) save-back — --save-as files the answer as a new synthesis/ page

Stack

No cloud services. No API keys. No data leaves your machine.

Requirements

• Python 3.11+
• Node.js 18+ (for QMD)
• Ollama with the qwen3:14b model pulled (~9.3GB)
• QMD (npm install -g @tobilu/qmd)
• Homebrew SQLite on macOS (brew install sqlite)
• ~15GB free disk space for models and embeddings
• ~12GB RAM recommended (16GB+ for comfort)
• Obsidian (optional but strongly recommended for browsing)

Tested on macOS (Apple Silicon, M3 Pro 18GB). Should work on Linux; Windows untested.

Installation

# Clone
git clone https://github.com/YOUR-USERNAME/llm-wiki.git
cd llm-wiki

# Create a virtual environment (uv is faster than pip, either works)
uv venv
source .venv/bin/activate
uv pip install -e .

# Pull the LLM (one-time, ~9.3GB)
ollama pull qwen3:14b

# Install QMD (the search backend)
npm install -g @tobilu/qmd

# Verify
wiki version
wiki --help

Quick start

# 1. Create a wiki in a folder of your choosing
mkdir my-wiki && cd my-wiki
wiki init

# 2. Drop some source documents in raw/, or use:
wiki add ~/Documents/papers --recursive

# 3. Run ingest (interactive by default — shows you entities/concepts
#    before filing, with a y/n prompt per source)
wiki ingest

# First query triggers QMD to download its embedding + reranker models
# (~2GB, one-time). Subsequent queries are fast.

# 4. Ask questions
wiki query "what are the main themes across these documents?"

# 5. Save a good answer as a synthesis page
wiki query "compare X vs Y" --save-as x-vs-y-comparison

# 6. Health-check and auto-fix
wiki lint --fix

# 7. Browse the vault in Obsidian
open wiki/   # then "Open folder as vault"

Commands

Run wiki <command> --help for full options on any command. See USAGE.md for a full walkthrough.

Example output

A real ingest against notes.txt (28 words about Qwen3):

Source #1  raw/notes.txt
  parsing…
  extracting entities and concepts (thinking mode)…

Title: Quick Notes on Qwen
Slug:  quick-notes-on-qwen

Summary:
  Qwen is a family of large language models developed by Alibaba Cloud.
  The latest version, Qwen3, introduces a thinking mode designed to enhance
  performance on complex reasoning tasks.

Entities (3):
  + alibaba-cloud (organization)  Alibaba Cloud
  + qwen (product)                Qwen
  + qwen3 (product)               Qwen3

Concepts (2):
  + large-language-models                 Large Language Models
  + thinking-mode-for-complex-reasoning   Thinking Mode for Complex Reasoning

File these? Will create/update ~6 wiki pages. [Y/n]: Y

created entity alibaba-cloud
created entity qwen
created entity qwen3
created concept large-language-models
created concept thinking-mode-for-complex-reasoning
created source  quick-notes-on-qwen

✓ Ingested Quick Notes on Qwen — 6 created, 0 updated

That's 6 cross-linked pages from a 28-word input, each with YAML frontmatter, [[wikilinks]] between them, and provenance back to the source. Open Obsidian's graph view and you'll see the cluster light up.

A real query against 11 ingested pages:

> wiki query "how does multi-head attention differ from self-attention?"

  searching wiki (BM25 + vector + rerank)…
  found 8 relevant page(s):
    1. 0.93 concepts/multi-head-attention.md      Multi-Head Attention
    2. 0.55 concepts/self-attention-mechanism.md  Self-Attention Mechanism
    3. 0.40 entities/attention-is-all-you-need.md Attention is All You Need
    ...

  synthesizing answer…

Multi-head attention and self-attention are related but distinct mechanisms:

1. **Scope and Parallelism**
   - Self-attention is a single mechanism where each position in the input
     computes attention weights based on all other positions
     [[concepts/self-attention-mechanism]].
   - Multi-head attention extends this by using multiple parallel attention
     heads, allowing the model to focus on diverse patterns simultaneously
     [[concepts/multi-head-attention]].

2. **Information Capture**
   - Self-attention focuses on a single representation.
   - Multi-head aggregates information from multiple heads, each capturing
     different aspects (syntactic vs semantic, etc.)
     [[concepts/multi-head-attention]].

[... etc.]

Every claim is cited. Every citation points to a page that actually exists.

Lint example

> wiki lint

╭─────────── Lint Report ────────────╮
│ Health score: 57/100               │
│ Pages checked: 12                  │
│                                    │
│   2 errors · 21 warnings · 0 infos │
╰────────────────────────────────────╯

──── Errors (2) ────

  synthesis/transformers-and-llms.md
    ✗ broken_wikilink: Broken wikilink: [[entities/introduction-to-transformers]]
      → Either create the page or remove the link.

──── Warnings (21) ────

  entities/qwen.md
    ! malformed_wikilink: 'sources/quick-notes-on-qwen.md' should be
      'sources/quick-notes-on-qwen'
      ✓ auto-fixable

  [... 20 more warnings ...]

> wiki lint --fix
✓ auto-fixed: 11

> wiki lint
╭─────────── Lint Report ───────────╮
│ Health score: 100/100             │
│   0 errors · 0 warnings · 0 infos │
╰───────────────────────────────────╯
✓ No issues found. Your wiki is in good shape!

Project status

Current version: v0.8.1 — production-ready for personal use.

Possible future work

• Hugging Face Spaces deployment (smaller model, API-compatible)
• Dashboard showing live active-job count
• Static HTML export for sharing the wiki
• Multi-user / team features
• Mobile-friendly web UI
• Fine-tuned query expansion model
• Confidence scoring per extracted claim
• OCR support for scanned PDFs
• EPUB support

Credits

• Andrej Karpathy — for the LLM-Wiki pattern described in this gist. This project is a direct implementation of the idea.
• QMD by Tobi Lütke — the hybrid search backend that does all the heavy lifting for query-time retrieval.
• Qwen3 by Alibaba Cloud — the local LLM doing the reading, writing, and synthesis.
• Ollama — the runtime that makes local LLM inference painless on Apple Silicon.
• Obsidian — saved me from writing my own graph view.

License

MIT — see LICENSE.

"The tedious part of maintaining a knowledge base is not the reading or the thinking — it's the bookkeeping. Humans abandon wikis because the maintenance burden grows faster than the value. LLMs don't get bored, don't forget to update a cross-reference, and can touch 15 files in one pass." — Karpathy