README.md

Efficient Data Conversation

Chat with your data while uploading a pdf file and using a local LLM.

Table of Contents

• PDF File Structure Support
• Language Support
• Key Dependencies
• Setup Guidelines
• System Support
• Credits

PDF File Structure Support:

1. Upcoming:Files with well organized tables i.e.: a single row/column ins not divided in multi row/column
2. Usually Research Paper Structure:
   • Abstract
   • Intorduction
   • Background Works
   • Dataset
   • Methodology
   • Result Analysis
   • Discussion
   • Future Works
   • Conclusion
3. No Image support for now
4. Up coming: meta data support

Language Support:

1. English
2. Others are loading...

Key Dependencies:

• Ollama with or without GPU
• Sentence-transformers
• Langchain

The models in use:

1. Attempted Sentence Embedding, chosen on mainly MTEB leaderboard and personal experience:
   • multi-qa-distilbert-cos-v1
   • multi-qa-mpnet-base-dot-v1
   • e5-base-v2
   • RobBERT [Currently, In Use]
2. Attemtped LLMs, chosen based on Mistral-7b's acceptable performence for low resource devices:
   • Mistral-7b: instruct-v0.2-q2_K
   • Mistral-7b: instruct-v0.2-q5_K_M
   • Mistral-7b: instruct-v0.2-q6_K [Currently, In use]

To store models, open a sub-directory inside the "api" directory open a directory.

For example: "lang_models":

Setup Guidelines:

1. OS tested: Ubuntu>=20.04 LTS
2. Create a Python>=3.11 environment using conda or virtual env
3. Use the requirements file to install the dependencies:

pip install -r requirements.txt

4. Use Ollama docker and Huggingface to pull/download all the models, refer to section: Key Dependencies for details and where to store the models inside your machine.
5. Set the .env file according to the .env.example structure. Note: For CPU inference, set USE_GPU=0
6. From the parent directory, to run the system, execute the command below in the termnal:

streamlit run api/app.py

System Support:

1. Integrated frontend with Streamlit
2. Up-coming: Separated backend support
3. Up-coming: Docker support

Credits and special thanks to my friends:

1. Sharif Ahamed, MSc. in AI, University of Bradford, Bradford, United Kingdom, Email:
   • For advising me through
2. Soroush Yaghoubi, BSc. In Informatics, Technical University Dortmund, Dortmund, Germany:
   • For the frontend idea and more works in future