This project provides a Chinese history question-answering application using Retrieval-Augmented Generation (RAG) with vector databases. It targets users seeking accurate historical information and aims to mitigate LLM hallucination by grounding responses in retrieved historical documents.

How It Works

The core approach leverages the LlamaIndex framework for RAG. It offers two primary deployment options: a local Milvus vector database setup using the BAAI/bge-base-zh-v1.5 embedding model, or a cloud-based Zilliz Cloud Pipelines service for document processing and retrieval. Both methods utilize OpenAI's GPT-4 for text generation. The system ingests historical texts, slices them, generates embeddings, and stores them in a vector index for efficient retrieval.

Quick Start & Requirements

• Installation: pip install -r requirements.txt (virtual environment recommended).
• Prerequisites: OpenAI API Key, Docker (for Milvus), Milvus 2.3.3 or Zilliz Cloud account, Python 3.
• Setup: Requires configuring API keys and potentially setting up a local Milvus instance or Zilliz Cloud account.
• Docs: Zhihu Article, Windows Instructions, Local LLM Support, Gemini Support, Gradio UI, Tongyi Qianwen Support.

Highlighted Details

• Supports local LLM services (e.g., fastchat) and Gemini models via a proxy.
• Includes a Gradio-based web UI for interactive use.
• Allows customization of embedding and reranker models via cfgs/config.yaml.
• Zilliz Cloud Pipelines offer scalability and managed RAG services.

Maintenance & Community

The project has seen recent updates (June 2024) to LlamaIndex and reranker functionality. Contributions are noted from users like darius-gs, BetterAndBetterII, leyiang, and taihaozesong.

Licensing & Compatibility

The repository does not explicitly state a license in the README. Compatibility for commercial use or closed-source linking is not specified.

Limitations & Caveats

The Zilliz Cloud Pipelines solution currently only supports importing documents via URL. Customizing LLMs requires modifying executor.py in addition to configuration files. The project relies on OpenAI's GPT-4 by default, and alternative LLM integration may require code changes.