awesome-llm-apps/rag_tutorials/local_hybrid_search_rag

🖥️ Local RAG App with Hybrid Search

A powerful document Q&A application that leverages Hybrid Search (RAG) and local LLMs for comprehensive answers. Built with RAGLite for robust document processing and retrieval, and Streamlit for an intuitive chat interface, this system combines document-specific knowledge with local LLM capabilities to deliver accurate and contextual responses.

Demo:

https://github.com/user-attachments/assets/375da089-1ab9-4bf4-b6f3-733f44e47403

Quick Start

For immediate testing, use these tested model configurations:

# LLM Model
bartowski/Llama-3.2-3B-Instruct-GGUF/Llama-3.2-3B-Instruct-Q4_K_M.gguf@4096

# Embedder Model
lm-kit/bge-m3-gguf/bge-m3-Q4_K_M.gguf@1024

These models offer a good balance of performance and resource usage, and have been verified to work well together even on a MacBook Air M2 with 8GB RAM.

Features

• Local LLM Integration:

  • Uses llama-cpp-python models for local inference
  • Supports various quantization formats (Q4_K_M recommended)
  • Configurable context window sizes

• Document Processing:

  • PDF document upload and processing
  • Automatic text chunking and embedding
  • Hybrid search combining semantic and keyword matching
  • Reranking for better context selection

• Multi-Model Integration:

  • Local LLM for text generation (e.g., Llama-3.2-3B-Instruct)
  • Local embeddings using BGE models
  • FlashRank for local reranking

Prerequisites

1. Install spaCy Model:

   pip install https://github.com/explosion/spacy-models/releases/download/xx_sent_ud_sm-3.7.0/xx_sent_ud_sm-3.7.0-py3-none-any.whl
   

2. Install Accelerated llama-cpp-python (Optional but recommended):

   # Configure installation variables
   LLAMA_CPP_PYTHON_VERSION=0.3.2
   PYTHON_VERSION=310 # 3.10, 3.11, 3.12
   ACCELERATOR=metal  # For Mac
   # ACCELERATOR=cu121  # For NVIDIA GPU
   PLATFORM=macosx_11_0_arm64  # For Mac
   # PLATFORM=linux_x86_64  # For Linux
   # PLATFORM=win_amd64  # For Windows
   
   # Install accelerated version
   pip install "https://github.com/abetlen/llama-cpp-python/releases/download/v$LLAMA_CPP_PYTHON_VERSION-$ACCELERATOR/llama_cpp_python-$LLAMA_CPP_PYTHON_VERSION-cp$PYTHON_VERSION-cp$PYTHON_VERSION-$PLATFORM.whl"
   

3. Install Dependencies:

   git clone https://github.com/Shubhamsaboo/awesome-llm-apps.git
   cd awesome-llm-apps/rag_tutorials/local_hybrid_search_rag
   pip install -r requirements.txt
   

Model Setup

RAGLite extends LiteLLM with support for llama.cpp models using llama-cpp-python. To select a llama.cpp model (e.g., from bartowski's collection), use a model identifier of the form "llama-cpp-python/<hugging_face_repo_id>/@<n_ctx>", where n_ctx is an optional parameter that specifies the context size of the model.

1. LLM Model Path Format:

   llama-cpp-python/<repo>/<model>/<filename>@<context_length>
   

   Example:

   bartowski/Llama-3.2-3B-Instruct-GGUF/Llama-3.2-3B-Instruct-Q4_K_M.gguf@4096
   

2. Embedder Model Path Format:

   llama-cpp-python/<repo>/<model>/<filename>@<dimension>
   

   Example:

   lm-kit/bge-m3-gguf/bge-m3-Q4_K_M.gguf@1024
   

Database Setup

The application supports multiple database backends:

• PostgreSQL (Recommended):
  • Create a free serverless PostgreSQL database at Neon in a few clicks
  • Get instant provisioning and scale-to-zero capability
  • Connection string format: postgresql://user:pass@ep-xyz.region.aws.neon.tech/dbname

How to Run

1. Start the Application:

   streamlit run local_main.py
   

2. Configure the Application:

   • Enter LLM model path
   • Enter embedder model path
   • Set database URL
   • Click "Save Configuration"

3. Upload Documents:

   • Upload PDF files through the interface
   • Wait for processing completion

4. Start Chatting:

   • Ask questions about your documents
   • Get responses using local LLM
   • Fallback to general knowledge when needed

Notes

• Context window size of 4096 is recommended for most use cases
• Q4_K_M quantization offers good balance of speed and quality
• BGE-M3 embedder with 1024 dimensions is optimal
• Local models require sufficient RAM and CPU/GPU resources
• Metal acceleration available for Mac, CUDA for NVIDIA GPUs

Contributing

Contributions are welcome! Please feel free to submit a Pull Request.