LangChain Simple RAG Pipeline:

This repo implements a simple RAG framework for a Pharmacy store chatbot. We use medical QA data from https://github.com/Toyhom/Chinese-medical-dialogue-data and a virtue inventory by web crawling from https://www.dayi.org.cn/ to build the knowledge database.

Medical QA data filtering

Data processing and ChatGLM3 QA demo in `test.ipynb

Step 0. Environment

Please first prepare an environment that can run ChatGLM3-6B. Then,

pip install langchain gradio chromadb sentence-transformers unstructured markdown

If you want to launch the system with our provided models, for simplicity, please unzip the models under the relative path ../autodl-tmp/model/

Step 1. Build VectorDB

cd RAG
conda activate nlp

```python

#### build VectorDB for demo QA files
python init_vectorDB.py --docs_dir ./docs/demo --persist_dir ./vectorDB/demo --embedder ../autodl-tmp/model/bge-m3

#### build VectorDB for a complete drug dictionary
python init_vectorDB.py --docs_dir ./docs/drug --persist_dir ./vectorDB/drug --embedder ../autodl-tmp/model/bge-m3 --chunk_size 100

#### build VectorDB for complete QA files
python init_vectorDB.py --docs_dir ./docs/QA --persist_dir ./vectorDB/QA --embedder ../autodl-tmp/model/bge-m3 --chunk_size 100

We use bge-m3 as the embedder. Please specify the link to your embedder models by --embedder.

Note that to build to vectorDB for the whole QA dataset, one can use our provided codebase in the report or need to download and process the QA data from https://github.com/Toyhom/Chinese-medical-dialogue-data

Step 2. Launch Web-UI

# ##  chat with demo QA FILES & drug FILES
python run_gradio.py --comparison_mode --verbose --QA_vectordb ./vectorDB/demo --drug_dict_vectordb ./vectorDB/drug --embedder ../autodl-tmp/model/bge-m3 --llm ../autodl-tmp/model/chatglm3-6b

# ##  chat with complete QA & drug FILES
python run_gradio_new.py  --verbose --QA_vectordb ./vectorDB/QA --drug_dict_vectordb ./vectorDB/drug --embedder ../autodl-tmp/model/bge-m3 --llm ../autodl-tmp/model/chatglm3-6b 

# ##  chat with complete QA & drug FILES & Enable context classification and BERT recommmandation
python run_gradio_new.py  --verbose --QA_vectordb ./vectorDB/QA --drug_dict_vectordb ./vectorDB/drug --embedder ../autodl-tmp/model/bge-m3 --llm ../autodl-tmp/model/chatglm3-6b --bert_recommand --context_cls

comparison_mode will output the response with and without RAG.
verbose enables the verbose mode of LangChain. The full prompt will be displayed in the terminal.
bert_recommand will also load and output the recommendation using BERT in the web GUI.
context_cls will load and output the medical context classification using BERT in the web GUI.
embedder dir of the embedder models from Huggingface. Change it to the dir of your local embedder.
llm dir of the language model from Huggingface. Change it to the dir of your local llm.
drug_dict_vectordb the vector database for the inventory information.
``

Note that to display the BERT context classification and recommendation, please unzip our pretrained models under the relative path ../autodl-tmp/model/

【测试案例】

头痛吃什么药？

今天星期几？

夏天适合哪里旅游？

你们这里有感冒药吗？

维生素和抗生素还有库存吗？

钙片有没有卖？

这里有卖百多邦软膏吗？

我要买后悔药！

导购员，给我来一盒奥美拉唑！

高血压会遗传吗？

这几天发烧头痛、四肢无力怎么办？

我脾虚胃寒，湿气重，有什么药品推荐？

Linnore / RAG-LangChain