The project, as part of my bachelor's thesis, intends to address some limitations of traditional LLMs, namely, knowledge cut-off and hallucinations. In this case study, I used GROQ, LangChain and Meta Llama3-8B to build the system, which consists of a workflow which verifies, grades and improves question answering generation. It can be applied to any structured dataframe or textual data (corpus); in my final work, it has been used on financial news data. More information is available on my Thesis paper.
- Clone the repository.
- Create a conda environment using the requirements.yaml file with this command
conda env create -f requirements.yaml
- Activate the requirements env:
conda activate rag_environment
The project has a very simple structure, but it is incomplete. Under the data folder, you should insert your docs (text, a csv, or any file format supported by LangChain parser) to process. Once embeddings and indeces are created, they should be stored in the 'Data' subfolder, so that they can be easily retrieved.
├───Agentic-RAG-Project <- Folder root
│ ├───Data <- knowledge base data used for RAG, along with indeces and embeddings.
│ ├───main.ipynb <- notebook containing code to execute
│ ├───requirements.yaml <- packages to install
│ └───README.md
In order to use and test the code, please simply run the different chunks in the notebook either by running main.ipynb or another python file you may want to use for orchestration. Note that file paths in the notebook have to be defined! If a csv is under the data folder, you may want to load the data using pandas, polars or huggingchat loaders. Below, a simple example (assuming main.ipynb does not change location):
df = pl.read_csv('../data/'your_file.csv')
The project is licensed under the MIT License.