Pizza Restaurant Review Q&A System

This project uses LangChain, Ollama, and Chroma to answer questions about pizza restaurants based on reviews in realistic_restaurant_reviews.csv. It leverages CUDA, cuDNN, and PyTorch for GPU acceleration (tested on RTX 3060, 12 GB VRAM) and monitors GPU usage with pynvml in Monitor_cuda.py.

Features

• Answers questions (e.g., “What’s the best pizza in town?”) using review data.
• GPU-accelerated with llama3.2:latest (2.0 GB) and mxbai-embed-large:latest (669 MB).
• Uses Chroma for vector search (top 5 reviews).
• Monitors VRAM with pynvml and nvidia-smi.
• Interactive CLI for questions.

Prerequisites

• Hardware: NVIDIA GPU (e.g., RTX 3060, 12 GB VRAM), 16 GB RAM.
• Software: Windows 10/11 (tested), Python 3.10+, NVIDIA driver 566.36+, CUDA 12.6/12.7, Ollama.
• Dataset: realistic_restaurant_reviews.csv with Title, Review, Rating, Date.

Setup Instructions

1. Install NVIDIA Drivers and CUDA:

   • Get NVIDIA driver from NVIDIA.
   • Verify: nvidia-smi (should show CUDA 12.6/12.7).
   • cuDNN is bundled with Py pytorch.

2. Install Python and Virtual Environment:

   python -m venv venv
   .\venv\Scripts\Activate.ps1  # Windows

3. Install Dependencies:

   pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu126
   pip install -r requirements.txt

   requirements.txt:

   langchain
   langchain-ollama
   langchain-chroma
   pandas
   pynvml
   

4. Install Ollama:

   • Download from ollama.ai.
   • Pull models:

     ollama pull llama3.2:latest
     ollama pull mxbai-embed-large:latest

   • Verify:

     ollama list

     Should show:

     NAME                          ID              SIZE      MODIFIED
     llama3.2:latest               a----------5    2.0 GB    Recently
     mxbai-embed-large:latest      4----------7    669 MB    Recently
     

5. Prepare Dataset:

   • Place realistic_restaurant_reviews.csv in project root.
   • Format:

     Title,Review,Rating,Date
     "Great Pizza","Crispy crust, fresh toppings!",5,"2023-10-01"
     

Project Structure

pizza-restaurant-review/
├── main.py                  # Runs Q&A system
├── vector.py                # Handles embeddings and vector database
├── Monitor_cuda.py          # Monitors GPU memory
├── requirements.txt         # Python dependencies
├── realistic_restaurant_reviews.csv  # Review dataset
├── chrome_langchain_db/     # Chroma database (auto-generated)
└── venv/                    # Virtual environment

How It Works

1. Data Loading (vector.py):
   • Reads CSV, combines Title and Review, creates Document objects with rating, date.
2. Embedding (vector.py):
   • Uses mxbai-embed-large:latest (669 MB, 1024 dimensions) for review embeddings.
   • Stores in Chroma (chrome_langchain_db), retrieves top 5 reviews.
3. Q&A (main.py):
   • Takes user question, retrieves reviews, uses llama3.2:latest (2.0 GB) to answer.
4. GPU Acceleration:
   • Ollama uses PyTorch with CUDA/cuDNN, ~4-5 GB VRAM.
5. Monitoring:
   • Monitor_cuda.py uses pynvml for GPU stats.
   • nvidia-smi shows ollama.exe/python.exe usage.

Running the Project

1. Create Environment:

   python -m venv venv

2. Activate Environment:

   .\venv\Scripts\Activate.ps1

3. Run Q&A:

   python main.py

   • Ask questions (e.g., “What’s the best pizza in town?”), type q to quit.
   • Example:

     -------------------------------
     Ask your question (q to quit): whats the best pizza in town
     Based on reviews, [Pizza Place] has the best pizza for its crispy crust.
     

4. Run GPU Monitor:

   python Monitor_cuda.py

   • Outputs:

     Total GPU memory: 12288.00 MB
     Free GPU memory: ~7292.00 MB
     Used GPU memory: ~4824.00 MB
     

Monitoring GPU Usage

• In-Script (Monitor_cuda.py):

  import pynvml
  try:
      pynvml.nvmlInit()
      handle = pynvml.nvmlDeviceGetHandleByIndex(0)
      mem_info = pynvml.nvmlDeviceGetMemoryInfo(handle)
      print(f"Total GPU memory: {mem_info.total / 1024**2:.2f} MB")
      print(f"Free GPU memory: {mem_info.free / 1024**2:.2f} MB")
      print(f"Used GPU memory: {mem_info.used / 1024**2:.2f} MB")
  except pynvml.NVMLError as e:
      print(f"NVML Error: {e}")
  finally:
      pynvml.nvmlShutdown()

• External (nvidia-smi):

  nvidia-smi

  • Look for ollama.exe/python.exe using ~4-5 GB VRAM.
  • Continuous:

    nvidia-smi --query --display=COMPUTE,MEMORY -l 2

Dependencies

• Python Packages (requirements.txt):
  • langchain
  • langchain-ollama
  • langchain-chroma
  • vector
  • pandas
  • pynvml
• PyTorch with CUDA:

  pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu126

• Ollama Models:
  • llama3.2:latest (a----------5, 2.0 GB)
  • mxbai-embed-large:latest (4----------7, 669 MB)
• NVIDIA Stack:
  • CUDA 12.6/12.7
  • cuDNN (bundled with PyTorch)
  • NVIDIA driver 566.36+

Dataset

• File: realistic_restaurant_reviews.csv
• Format: CSV with:
  • Title: Review title
  • Review: Review text
  • Rating: 1-5
  • Date: e.g., “2023-10-01”
• Usage: Loaded by vector.py for embeddings.