GangDan (纲担)

LLM-powered knowledge management and teaching assistant with offline support.

GangDan (纲担) — Principled and Accountable.

Overview

GangDan is a local-first, offline programming assistant powered by Ollama and ChromaDB. It combines RAG-based knowledge management with teaching assistance tools, all running entirely on your machine — no cloud APIs required.

Features

Knowledge Management

• Unified Literature Search — Search arXiv, bioRxiv, medRxiv, Semantic Scholar, CrossRef, OpenAlex, DBLP, PubMed, and GitHub in one interface. AI-powered query refinement with automatic translation and synonym expansion.
• Batch Operations — Multi-select, select-all, batch convert (PDF/HTML/TeX to Markdown with image and formula preservation), batch add to knowledge base. Sort by relevance, date, or title.
• Smart Renaming — Downloaded papers automatically renamed to citation format: Author et al. (Year) - Title.pdf
• LLM-Generated Wiki — Build structured wiki pages from knowledge base content with cross-KB concept linking. Like Wikipedia for your documents.
• Image Gallery — Browse and search images stored in knowledge bases with context and source attribution.
• Document Manager — One-click download and indexing of 30+ library docs (Python, Rust, Go, JS, CUDA, Docker, etc.). Upload custom docs, batch operations, GitHub repo search, web search to KB.
• Custom Knowledge Base Upload — Upload your own Markdown (.md) and plain text (.txt) documents to create named knowledge bases with automatic indexing.

Teaching Assistant

• Question Generator — MCQ, short answer, fill-in-the-blank, true/false from KB content.
• Guided Learning — Auto-extract knowledge points, generate interactive lessons with Q&A.
• Deep Research — Multi-phase research pipeline: topic decomposition → RAG research → comprehensive reports.
• Lecture Maker — Generate structured lecture content from KB materials.
• Exam Generator — Create complete exam papers with answer keys from KB content.
• Literature Review & Paper Writer — Generate academic reviews and papers from KB content.

Core Features

• RAG Chat — Streaming chat with knowledge base retrieval and web search. Strict KB mode ensures grounded answers.
• Cross-Lingual Search — Automatically detects query and document languages, enabling cross-lingual RAG (e.g., query English documents in Chinese).
• Citation References — Each response automatically includes source document references for verification.
• AI Command Assistant — Natural language → shell commands, draggable to terminal.
• Built-in Terminal — Run commands with stdout/stderr display directly in the browser.
• Conversation Save/Load — JSON export/import for session continuity.
• 10-Language UI — Chinese, English, Japanese, French, Russian, German, Italian, Spanish, Portuguese, Korean.
• Dark/Light Theme — Full theme support with CSS variables.
• Offline by Design — Runs entirely on your machine. No cloud APIs required.

Multi-Provider LLM Support

GangDan supports a separated mode: local Ollama for chat/embedding/reranking, with optional external LLM providers for deep research and paper writing.

Provider	API Type	Use Case
Ollama (local)	ollama	Chat, Embedding, Reranking
DashScope	OpenAI-compatible	Deep Research, Paper Writing
MiniMax	OpenAI-compatible	Deep Research
Bailian Coding	Anthropic-compatible	Deep Research
OpenAI / DeepSeek / Moonshot	OpenAI-compatible	Deep Research
Custom	OpenAI-compatible	Any compatible API

CLI

• Streaming chat (gangdan chat "question"), interactive REPL (gangdan cli)
• KB operations, doc management, config, conversation persistence
• AI command generation, shell execution with safety checks
• Rich terminal output with formatted tables and syntax highlighting

Screenshots

Chat	Terminal

Documentation	Settings

Upload Documents	KB Scope Selection

Strict KB Chat with Citations

The above screenshot demonstrates Strict KB Mode in action: after selecting a specific knowledge base, the system retrieves content only from that KB and automatically appends a reference list at the end of each response, citing the source documents.

Load Conversation	Conversation Loaded

Save your chat as a JSON file and load it anytime to continue the conversation.

RAG Pipeline

The complete pipeline from document ingestion to retrieval:

1. Document Ingestion — Download from GitHub repositories or upload custom files (.rst, .py, .html, .cpp, .md)
2. Format Conversion — Automatic conversion to unified Markdown format
3. Sliding Window Chunking — Fixed-size segmentation with configurable overlap (default: 800 chars, 150 overlap)
4. Vector Embedding — nomic-embed-text model via Ollama API (768-dim vectors, 500-char truncation)
5. Vector Storage — ChromaDB with HNSW indexing and cosine similarity
6. Query Retrieval — Top-K search with distance filtering (threshold 1.5), deduplication, and context construction

Chunking Strategy

The sliding window approach ensures contextual continuity across chunk boundaries. Key parameters:

Parameter	Default	Range	Description
CHUNK_SIZE	800 chars	100-2000	Characters per chunk
CHUNK_OVERLAP	150 chars	N/A	Overlap between consecutive chunks
MIN_CHUNK	50 chars	N/A	Minimum chunk length threshold

Requirements

• Python 3.10+
• Ollama running locally (default http://localhost:11434)
• Chat model (e.g. ollama pull qwen2.5)
• Embedding model (e.g. ollama pull nomic-embed-text)

Installation

Method 1: Install from PyPI (Recommended)

pip install gangdan
gangdan                    # Web GUI
gangdan cli                # Interactive CLI
gangdan --port 8080        # Custom port

Method 2: Install from Source

git clone https://github.com/cycleuser/GangDan.git
cd GangDan
pip install -e .
gangdan

Open http://127.0.0.1:5000 in your browser.

Ollama Setup

ollama serve
ollama pull qwen2.5
ollama pull nomic-embed-text

Project Structure

GangDan/
├── pyproject.toml
├── README.md / README_CN.md
├── gangdan/
│   ├── __init__.py / __main__.py
│   ├── cli.py / cli_app.py          # CLI entry + REPL
│   ├── app.py                       # Flask backend
│   ├── learning_routes.py           # Learning module blueprint
│   ├── preprint_routes.py           # Preprint search + convert
│   ├── research_routes.py           # Paper search
│   ├── kb_routes.py                 # Custom KB management
│   ├── export_routes.py             # Export API
│   ├── core/                        # Shared modules
│   │   ├── config.py                # Config, i18n, translations
│   │   ├── ollama_client.py         # Ollama API
│   │   ├── chroma_manager.py        # ChromaDB
│   │   ├── vector_db.py             # Multi-backend vector DB
│   │   ├── kb_manager.py            # Custom KB CRUD
│   │   ├── conversation.py          # Chat history
│   │   ├── doc_manager.py           # Doc download/index
│   │   ├── wiki_builder.py          # LLM wiki generation
│   │   ├── preprint_fetcher.py      # Preprint search
│   │   ├── preprint_converter.py    # HTML/TeX/PDF → MD
│   │   ├── pdf_converter.py         # PDF → MD (marker/mineru/docling)
│   │   ├── export_manager.py        # Batch convert/export
│   │   ├── web_searcher.py          # Web search
│   │   └── ...
│   ├── templates/index.html         # Main SPA template
│   └── static/{css,js}/             # Frontend assets
├── tests/                           # Test suite
├── images/                          # Screenshots
├── diagrams/                        # Architecture diagrams (SVG)
└── removed/                         # Deprecated files

Configuration

All settings through the Settings tab: Ollama URL, chat/embedding/reranker models, proxy, context length, output language, vector DB type, LLM provider selection, and API keys.

Testing

pip install pytest pytest-cov
pytest tests/ -v
pytest tests/ --cov=gangdan

Academic Paper

For a detailed empirical study of the RAG pipeline and chunking strategies, see Article.md / Article_CN.md.

License

GPL-3.0-or-later. See LICENSE for details.