Th e codebase of C-RAG structures it contains 3 RAG Pipelines:
- Classic C-RAG
- Self C-RAG
- Adaptive RAG
- This project uses conda as an environment manager. The user must have Anaconda or Miniconda installed.
- This project uses git-hooks to check code quality when creating commits.
- Install conda (miniconda version): https://docs.conda.io/projects/conda/en/latest/user-guide/install/index.html
- Install conda (anaconda navigator version): https://anaconda.org/anaconda/conda
- Make sure conda directory (
C:\Users\<USER>\AppData\Local\miniconda3) or (C:\Users\<USER>\AppData\Local\conda) is added to PATH environment variable in Windows - Create conda environment from
env.yaml:conda env create -f env.yaml - Activate environment:
conda activate CRAG
- Create
.envfile in the project's root directory (based on .env-sample file). The content of.envis not stored in Git repository, because it contains secrets.
- Edit Run/Debug configurations in PyCharm and make sure to select the correct
.envfile and conda environment - In PyCharm choose
File -> Settings -> Python interpreterand selectCRAGenvironment
Before started using main.py you have to use ingestion.py which create vector store based on three links. Of course, you can use your own links as well.
main.py - It's based on basic streamlit library so to run app you have to run command streamlit run main.py
- C-RAG - the rag which grade documents if they relevant them use them, or it is not then use websearch (tavily to find additional information).
- Self C-RAG - similar to previous one but self check if info from websearch was useful and model does hallucinate if yes then repeat until he won't get accept that it use relevant info and did not hallucinate.
- Adaptive Self C-RAG - devlop on first check so If question is about something about LLM then check document otherwise it will pass this process and use only websearch.
In repo is langgraph.json which is compatible with LangGraph Studio: https://blog.langchain.dev/langgraph-studio-the-first-agent-ide/