Skip to content

adev175/llm-wiki

Repository files navigation

LLM Wiki

An Obsidian vault + MCP server that turns Claude into a personal knowledge compiler — it reads your research, extracts concepts, links ideas together, and answers questions from what it has learned.


Install

pip install -r mcp_server/requirements.txt

Add to %APPDATA%\Claude\claude_desktop_config.json:

{
  "mcpServers": {
    "llm-wiki": {
      "command": "python",
      "args": ["C:\\Users\\nhata\\PycharmProjects\\llm-wiki\\mcp_server\\server.py"]
    }
  }
}

Restart Claude Desktop. Done.


How to use

Just talk to Claude naturally:

  • "Here's an article about momentum trading — add it to the wiki" → Claude extracts concepts, creates/updates pages, links related ideas
  • "What do we know about mean reversion?" → Claude searches the wiki and answers with citations
  • "Find papers on reinforcement learning for trading" → Claude searches Arxiv, fetches relevant papers, saves them

Everything gets saved automatically. You don't need to learn any commands.


Features

Knowledge pages

Each topic gets its own Markdown file in vault/wiki/ with structured frontmatter. Pages are linked with [[wikilinks]] so nothing is an island. Claude never creates a page without first checking if one already exists.

Page types (determined by slug prefix):

Type Prefix What it stores
Concept concept- A technical idea or theory — definition, how it works, edge cases
Strategy strategy- A trading strategy — setup, signals, risk rules, empirical notes
Source source- Summary of a book, paper, or article — key claims, quotes, context
Entity entity- A person, firm, or tool — background, notable work, opinions
Decision decision- Why you chose approach X over Y — alternatives considered, reversibility
Log log- Captured conversation insight or daily research note
Paper paper- Arxiv paper — abstract, methods, results, links to related concepts

Synonym-aware search

Pages have an aliases field listing alternate names for the same concept:

aliases: ["mean reversion", "stat arb", "pairs trading"]

Searching for any of these terms finds the page — even if the query term never appears in the body text. This matters in quant where "momentum" = "trend following" = "time-series momentum" depending on the author.

Inbox for quick capture

Don't want to think about where something belongs right now? Claude drops it in vault/inbox/ with a timestamp. You can process it later, or ask Claude to sort through the inbox for you.

Daily research log

Every session can anchor to today's date via vault/wiki/log-YYYY-MM-DD.md. Useful for tracking what you read, what questions came up, and what decisions were made on a given day.

Arxiv integration

Claude can search Arxiv by topic and fetch full paper metadata in one step. Papers are saved to vault/papers/ and automatically linked to related concept pages.

Auto-timestamp hook

Every time a wiki page is edited, the updated field in frontmatter is updated automatically — no need to remember. Requires merging hooks/hook-configs.json into ~/.claude/settings.json.

Semantic search (optional)

For fuzzy conceptual search beyond keyword matching:

npm install -g @tobilu/qmd
qmd collection add ./vault --name llm-wiki
qmd context add llm-wiki "Quant trading research vault"
qmd embed

Once set up, the index rebuilds automatically after each Claude session.

Vault health check

python hooks/validate-frontmatter.py vault/wiki

Checks all pages for missing required fields, invalid types, malformed slugs, and non-list aliases.


Vault layout

vault/
├── inbox/     # Unprocessed quick-captures
├── raw/       # Original source text — never edited
├── wiki/      # All knowledge pages
├── papers/    # Arxiv papers
└── log.md     # Activity history

See CLAUDE.md for full agent conventions and workflow details.

About

implementation of karpathy's ideas about llm-wiki

Topics

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors

Languages