Local ChatPDF with DeepSeek R1

ChatPDF is a Retrieval-Augmented Generation (RAG) application that allows users to upload PDF documents and interact with them through a chatbot interface. The system uses advanced embedding models and a local vector store for efficient and accurate question-answering.

Features

• PDF Upload: Upload one or multiple PDF documents to enable question-answering across their combined content.
• RAG Workflow: Combines retrieval and generation for high-quality responses.
• Customizable Retrieval: Adjust the number of retrieved results (k) and similarity threshold to fine-tune performance.
• Memory Management: Easily clear vector store and retrievers to reset the system.
• Streamlit Interface: A user-friendly web application for seamless interaction.

---

Installation

Follow the steps below to set up and run the application:

1. Clone the Repository

git clone https://github.com/paquino11/chatpdf-rag-deepseek-r1.git
cd chatpdf-rag-deepseek-r1

2. Create a Virtual Environment

python3 -m venv venv
source venv/bin/activate

3. Install Dependencies

Install the required Python packages:

pip install -r requirements.txt

Make sure to include the following packages in your requirements.txt:

streamlit
langchain
langchain_ollama
langchain_community
streamlit-chat
pypdf
chromadb

4. Pull Required Models for Ollama

To use the specified embedding and LLM models (mxbai-embed-large and deepseek-r1), download them via the ollama CLI:

ollama pull mxbai-embed-large
ollama pull deepseek-r1:latest

---

Usage

1. Start the Application

Run the Streamlit app:

streamlit run app.py

2. Upload Documents

• Navigate to the Upload a Document section in the web interface.
• Upload one or multiple PDF files to process their content.
• Each file will be ingested automatically and confirmation messages will show processing time.

3. Ask Questions

• Type your question in the chat input box and press Enter.
• Adjust retrieval settings (k and similarity threshold) in the Settings section for better responses.

4. Clear Chat and Reset

• Use the Clear Chat button to reset the chat interface.
• Clearing the chat also resets the vector store and retriever.

---

Project Structure

.
├── app.py                  # Streamlit app for the user interface
├── rag.py                  # Core RAG logic for PDF ingestion and question-answering
├── requirements.txt        # List of required Python dependencies
├── chroma_db/              # Local persistent vector store (auto-generated)
└── README.md               # Project documentation

---

Configuration

You can modify the following parameters in rag.py to suit your needs:

1. Models:

   • Default LLM: deepseek-r1:latest (7B parameters)
   • Default Embedding: mxbai-embed-large (1024 dimensions)
   • Change these in the ChatPDF class constructor or when initializing the class
   • Any Ollama-compatible model can be used by updating the llm_model parameter

2. Chunking Parameters:

   • chunk_size=1024 and chunk_overlap=100
   • Adjust for larger or smaller document splits

3. Retrieval Settings:

   • Adjust k (number of retrieved results) and score_threshold in ask() to control the quality of retrieval.

---

Requirements

• Python: 3.8+
• Streamlit: Web framework for the user interface.
• Ollama: For embedding and LLM models.
• LangChain: Core framework for RAG.
• PyPDF: For PDF document processing.
• ChromaDB: Vector store for document embeddings.

---

Troubleshooting

Common Issues

1. Missing Models:

   • Ensure you've pulled the required models using ollama pull.

2. Vector Store Errors:

   • Delete the chroma_db/ directory if you encounter dimensionality errors:

     rm -rf chroma_db/

3. Streamlit Not Launching:

   • Verify dependencies are installed correctly using pip install -r requirements.txt.

---

Future Enhancements

• Memory Integration: Add persistent memory to maintain conversational context across sessions.
• Advanced Analytics: Include visual insights from retrieved content.
• Expanded Model Support: Support additional embedding and LLM providers.

---

License

This project is licensed under the MIT License. See the LICENSE file for details.

---

Acknowledgments

• LangChain
• Streamlit
• Ollama