Document-RAGx

Purpose

This project aims to provide the end-to-end interface for RAG from documents.

Version 1.0

Structure

Frameworks & Tools

- LLM is self-hosted using ollama: llama3.2  

llama3.2 Model Link

- Python 3.11.9 

- LangChain 0.3.26 

- instructor 1.9.0

- chainlit 2.6.0

Main Engine

Document-RAGx aims to provide universal interface for hybrid retriever.

- Preprocessing Pipeline 

    1. A list of documents is parsed using PyMuPDF implemented in LangChain. 
    2. Parsed documents are splitted using RecursiveCharacterTextSplitter.
    3. Keywords are extracted first from documents using structured-output-generation by instructor leveraging llama3.2. 
    4. Extracted keywords and embeddings are stored in Chroma Vector DB. 


- Retrieval & Generation Pipeline 

    1. Keywords are extracted from user's query. 
    2. Extracted keywords and embedding of the query is used for hybrid retrieval. 
    3. Retrieved result is used for further generation. 

Front-End & Back-End

Chat interface is implemented using chainlit

How to use

Preprocess

• Run python preprocess.py --data_dir ./data --strategy naive --mode new

  • data_dir could be an arbitrary directory containing PDF files.

  • strategy should be naive. Currently only naive strategy (simple hybrid retrieval & generation) is implemented.

  • mode could be new or add.

    In new mode generates a new snapshot.json that represents all the documents inside the data directory and creates knowledge base.

    In add mode determines the new files based on the original snapshot.json and add additional documents to knowledge base.

Execute

• Activate virtual environment .venv\Scripts\activate

  • Make sure you installed required packages

• chainlit run app.py

• Visit http://localhost:8000

PS

If you have any issues, please visit here and leave me a message! :-)