Enterprise-grade Next-generation Intelligent Knowledge OS: Integration of Deep Research, Knowledge Graphs, and Hybrid RAG

English | 简体中文

In the era of enterprise knowledge management, we face the dual challenges of "Data Silos" and "Knowledge Hallucination". Traditional RAG (Retrieval-Augmented Generation) is often limited by shallow semantic matching and fails to understand the deep topological relationships between pieces of knowledge.

KnowledgeOS aims to break the deadlock of "dead documents, unsearchable data, and lack of correlation." We innovatively combine the deep interaction philosophy of LLM-Wiki with the zero-hallucination static analysis of Graphify to build a dynamically growing, reasoning-capable enterprise knowledge brain. It is not just a repository, but a self-evolving intelligent operating system.

Designed for B2B commercial scenarios. Based on a rigorous authentication system for Tenants and Users, it achieves complete physical isolation of data at the Project-Level.

• Physical Isolation: Documents, vector indices, and knowledge graphs of different projects are completely independent at the storage layer.
• Security & Sovereignty: Ensures data sovereignty between different departments and project teams within an enterprise.

The system automatically adopts the most suitable processing track for different data streams:

• Static Analysis Track (Code/Structured): Utilizes Tree-sitter (AST) to statically parse code and structured data with 0 Token consumption. It extracts definitions and references with precision, completely eliminating LLM hallucinations.
• Semantic Reasoning Track (Unstructured): For PDFs, Word docs, and web pages, it employs a 2-Step Chain-of-Thought (CoT): first analyzing conceptual conflicts, then generating logical relations, transforming dead text into a living knowledge web.
• Multi-Media Track (Audio/Video): YouTube and Bilibili videos are transcribed via Faster-Whisper + yt-dlp, then fed into the semantic pipeline for deep knowledge extraction.

One-click content extraction from anywhere on the web via a three-level intelligent fetch chain:

• Supported Platforms: YouTube, Bilibili, Zhihu, WeChat Official Accounts, X/Twitter, arXiv, and any web page.
• Three-Level Fetch Chain: trafilatura (static extraction) → Jina Reader (cloud Playwright with JS rendering) → Tavily Extract (backup).
• Cookie Management: Centrally manage platform authentication cookies via the Admin Web dashboard (supports Header String and Netscape formats).
• Source Traceability: Original URLs are saved to the database, enabling one-click browser access to the source material.

Full Ingest Pipeline: From Raw Input to Structured Knowledge Node

Integrated with LiteLLM gateway to balance cost and intelligence:

• Smart Dispatching: Step 1 (coarse analysis) uses cost-effective models (e.g., DeepSeek), while Step 2 (deep graph reasoning) utilizes top-tier models (e.g., Gemini 1.5 Pro / Claude 3.5).
• Privacy Mode: Supports locally deployed Ollama, ensuring core confidential data never leaves your network.

A built-in Agent with "curiosity." When the LLM identifies knowledge gaps during document analysis, it automatically triggers deep research:

• Autonomous Exploration: Automatically executes web crawling to compare the latest industry information.
• Persistence of History: All research reports are saved as Markdown, with real-time status streaming in the UI, and allow for secondary editing and ingestion at any time.

Autonomous Deep Search and Real-time Ingestion Interface

Drawing inspiration from the rigor of the financial industry, we introduce a Maker-Checker mechanism.

• Quality Gate: Knowledge nodes generated by LLMs must pass through an expert Audit Pipeline before being injected into the enterprise public vector pool.
• Anti-Explosion: Effectively filters out low-quality, duplicate, or conflicting information to maintain the absolute purity of the knowledge base.

• Smooth Experience: Based on Sigma.js and graph pruning algorithms, achieving lightning-fast zooming and searching even with tens of thousands of nodes.
• Real-time Transparency: Uses Redis + SSE protocols to stream every step of the LLM's thinking, crawling, and parsing status to the frontend UI in real-time, eliminating "black box" waiting.

Real-time Knowledge Graph Topology Rendering

• Core Framework: Python 3.12 + FastAPI + Celery (Distributed Task Queue)
• Data Storage:
  • PostgreSQL: Relational data, user status, and cookie management.
  • LanceDB: Next-gen Serverless vector database for extreme project isolation.
  • Redis: Task scheduling and SSE log relay.
• Content Extraction:
  • trafilatura + Faster-Whisper + yt-dlp for local pipeline processing.
  • Jina Reader + Tavily Extract for cloud-based JS rendering fallback.
• Deployment: Fully containerized, supporting Docker Compose one-click startup.

• Stack: Tauri v2 + React 19 + Vite + Tailwind CSS
• Visualization: Sigma.js + Graphology

Full System Architecture Overview

Complete Technology Stack: Every Layer, Every Technology

• API Documentation (v0.5.6)
• Changelog
• Multi-Platform URL Ingestion Plan (Chinese)

Please refer to the Deployment Guide for configuration.

cd server
cp .env.example .env  # Fill in your API Keys
docker compose up -d
docker exec -it kos_api python init_db.py

Access http://localhost:8080/admin to manage:

• Model Configuration: Set API keys and model routing (Chat, Wiki Engine, Translator)
• Platform Cookies: Manage authentication cookies for YouTube, Zhihu, Bilibili etc.
• Shared Projects: Publish knowledge base packages and manage tenant access

cd client
npm install
npm run tauri dev

The development of KnowledgeOS is deeply inspired by the open-source community. Special thanks to the following projects for their inspiration and code contributions:

• LLM-Wiki: Provided the core CoT logic for knowledge extraction and the frontend interaction architecture.
• Graphify: Provided high-performance static code analysis and structured definition references.

This project is licensed under the GNU General Public License v3.0.

KnowledgeOS - Guarding Data Sovereignty, Building Private Brains