multi-agent-wiki-llm

A self-maintaining AI/ML research wiki driven by autonomous LLM agents. Drop in papers, articles, or notes — the agents read them, extract the entities and concepts, write structured Markdown pages, cross-link them, and keep the whole thing consistent. Then chat with it.

Inspired by Andrej Karpathy's wiki concept gist.

🧩 Architecture

Why this shape? The agent loop in base_agent.py is intentionally generic — it streams AgentEvents as it calls tools, regardless of which agent is running. Swapping models is a one-line change in router.py. Adding a tool is "write the function → add the schema → add to the dispatch table." Three places, no glue.

🛠️ Tech Stack

Backend · Python 3.10+ · FastAPI · google-genai SDK · pydantic-settings · pypdf · SQLite (chat history)

Frontend · React 19 · Vite · Tailwind CSS v4 (@theme) · react-force-graph-2d · react-markdown + remark-math + rehype-katex · native WebSocket

Storage · Plain Markdown (data/wiki/) · SQLite (data/chats.db) · raw uploads (data/raw/)

🚀 Quick Start

Prerequisites

• Python 3.10+
• Node.js 18+
• A free Google AI Studio API key

Run

Three terminals.

Note: --reload-dir backend is required. Without it, every file you upload to data/ triggers a server reload and drops the live WebSocket mid-ingest.