This project provides a 100% free, private, and locally runnable RAG chatbot powered by DeepSeek-7B. It targets users who want to build sophisticated chatbots from their own documents (PDF, DOCX, TXT) without internet connectivity, offering enhanced accuracy and contextual understanding through advanced RAG techniques.

How It Works

The chatbot employs a hybrid retrieval strategy combining BM25 and FAISS for initial document chunk selection. It then integrates GraphRAG to build a knowledge graph from the documents, enabling deeper relational understanding. A neural reranker (Cross-Encoder) further refines the relevance of retrieved chunks. Query expansion via HyDE and chat history integration improve context and coherence, with DeepSeek-7B generating the final responses.

Quick Start & Requirements

• Installation: Clone the repository and install dependencies (pip install -r requirements.txt) or use Docker.
• Prerequisites: Ollama must be installed and running to pull and serve models (ollama pull deepseek-r1:7b, ollama pull nomic-embed-text). Requires Python 3.x.
• Running: Execute streamlit run app.py after setting up Ollama. Docker options are also provided.
• Docs: https://github.com/SaiAkhil066/DeepSeek-RAG-Chatbot

Highlighted Details

• Hybrid retrieval (BM25 + FAISS) augmented with GraphRAG for enhanced contextual understanding.
• Neural reranking using a Cross-Encoder model for improved relevance.
• HyDE (Hypothetical Document Embeddings) for query expansion.
• Chat history integration for maintaining conversational context.

Maintenance & Community

The project encourages community contributions via pull requests and issues. Feedback and discussions are welcomed on Reddit.

Licensing & Compatibility

The project is 100% free. Specific licensing details for the code and dependencies are not explicitly stated in the README, but the emphasis on "free" and "private" suggests a permissive stance for local use.

Limitations & Caveats

The README mentions upcoming features like model selection via UI and suggested questions, indicating the project is actively under development. Specific performance benchmarks or hardware requirements beyond running Ollama are not detailed.