RAG system
Using LLM and RAG to build a Q-A system based on specific knowledge. GUI is implemented by Streamlit.
Author: Kevin Stark, Jiarui Feng, Bowen Wang
Date: 2024/07/07
Version: 1.4
conda create -n rag-system python=3.10
conda activate rag-system
pip install streamlit, PyPDF2, sentence_transformers, faiss-cpu, requests, langchain, dashscope # dashscope for qianwenPut the files to be read under ./documents/ directory manually, or upload them in GUI.
Run it by streamlit run app.py (Using default model: gpt)
Run it by streamlit run app.py qwen (Using model: qian wen)
.
├── app.py
├── llmapi.py
├── cache
│ ├── cache_of_documents # Cache of loaded files
│ │ └── cache_of_documents.json
│ └── cache_of_embeddings # Cache of embeddings files
│ │ ├── cache_of_embeddings_batch_1.npy
│ │ ├── cache_of_embeddings_batch_2.npy
│ │ ├── etc.
│ └── cache_of_embeddings.npy
├── documents # Put the files manually here, or upload them in GUI.
│ ├── documents_1.pdf
│ └── etc.
└── README.md
v1.4:感谢@EPTansuo 的PR，增加Qwen LLM的支持。修正README的错误，faiss的库安装需要pip install faiss-cpu或者pip install faiss-gpu，而不是单独安装faiss。
v1.3: 更新缓存检查机制和命名规范；调整GUI，增加侧边栏，因为随着交互内容增加，想上传新文件/查看已上传文件信息要翻很久，如果有侧边栏/冻结首栏比较好。embeddings的缓存增量更新还没做。
v1.2: 添加上传文件功能。
v1.1: 添加多个注释方便阅读；优化函数preprocess_documents名改为split_text_chunks；在GUI里增加各步骤执行成功的反馈；考虑添加local llm；Streamlit里对'''多行注释'''好像有些独特用法，st.markdown (TD)
v1.0: 增加文档载入和向量嵌入的缓存机制；添加向量嵌入的批处理操作避免内存占用过大；暂时没有添加“上传文档”按钮，没有做文件更新缓存变动检查。