Adding GTC 2025 DLIT71491 GraphRAG notebooks (#288)

* Adding GTC 2025 DLIT71491 GraphRAG notebooks with secure configuration

* Adding GTC 2025 DLIT71491 GraphRAG notebooks with secure configuration

Author: rohrao

community/knowledge_graph_rag/GTC25_DLI/.dockerignore

@@ -0,0 +1,4 @@
+.git
+**/.ipynb_checkpoints
+# **/data
+**/dask-worker-space

community/knowledge_graph_rag/GTC25_DLI/.env

@@ -0,0 +1,5 @@
+COMPOSE_PROJECT_NAME=gtc25_kgrag_dli
+DEV_NGINX_PORT=9999
+NGC_API_KEY=#
+NVIDIA_API_KEY=#
+HUGGINGFACE_TOKEN=#

community/knowledge_graph_rag/GTC25_DLI/.gitignore

@@ -0,0 +1,3 @@
+**/.ipynb_checkpoints
+# **/data
+**/dask-worker-space

community/knowledge_graph_rag/GTC25_DLI/.launchignore

@@ -0,0 +1,17 @@
+# This file defines paths that should not be included in the
+# launch tarball.
+# This includes any paths in the docker container.
+# If a dataset loader is used, this file should also include
+# the paths to any datasets.
+Dockerfile
+.launchignore
+.dockerignore
+.gitignore
+vm_properties.json
+entrypoint.sh
+assessment/*
+.git
+**/.ipynb_checkpoints
+# **/data
+**/dask-worker-space
+**/mydask*

community/knowledge_graph_rag/GTC25_DLI/Dockerfile

@@ -0,0 +1,24 @@
+# Use NVIDIA's AI Workbench base image
+FROM nvcr.io/nvidia/ai-workbench/python-cuda120:1.0.3
+
+# Set the working directory in the container
+WORKDIR ./
+
+# Copy the application code into the container (optional)
+# COPY ./app_code ./app
+
+# Update and install additional system dependencies (if needed)
+RUN apt-get update && apt-get install -y --no-install-recommends \
+    build-essential \
+    curl  \
+    wget \
+    && apt-get clean && rm -rf /var/lib/apt/lists/* 
+
+# Install Python dependencies
+COPY requirements.txt ./requirements.txt
+RUN pip install --no-cache-dir -r requirements.txt
+
+EXPOSE 8888
+
+# ADD entrypoint.sh /usr/local/bin
+# ENTRYPOINT ["/usr/local/bin/entrypoint.sh"]

community/knowledge_graph_rag/GTC25_DLI/README.md

@@ -0,0 +1,80 @@
+# Knowledge Graph-based RAG System
+
+This repository contains the materials for the NVIDIA GTC 2025 Deep Learning Institute course: "Structure From Chaos: Accelerate GraphRAG With cuGraph and NVIDIA NIM" [DLIT71491].
+
+You can access the online course video at: [NVIDIA On-Demand](https://www.nvidia.com/en-us/on-demand/session/gtc25-dlit71491/)
+
+## Overview
+
+This course demonstrates how to build a Knowledge Graph-based RAG system for financial documents, specifically SEC 10-K filings. The system extracts structured knowledge triplets, builds a dynamic knowledge graph, and provides enhanced query capabilities.
+
+You'll learn how to integrate large language models (LLMs) with NVIDIA Inference Microservices (NIM) and cuGraph to create cutting-edge, graph-based AI solutions for handling complex, interconnected data. The course covers fine-tuning techniques, Langchain agents, and GPU-accelerated graph analytics to enhance AI capabilities and retrieval-augmented generation (RAG) evaluation.
+
+## Prerequisites
+
+- Docker and docker-compose
+- 4x NVIDIA A100 GPUs required for the LLM fine-tuning workflow (Notebook 2)
+- NVIDIA API key (sign up at [NVIDIA AI portal](https://developer.nvidia.com/))
+
+## Setup
+
+1. Clone this repository
+2. Set your NVIDIA API key in the `.env` file:
+   ```
+   NVIDIA_API_KEY=your-nvapi-key
+   NGC_API_KEY=your-ngc-key  # For some containerized models
+   ```
+3. Start the containers:
+   ```bash
+   docker-compose up -d
+   ```
+4. Access the Jupyter Lab environment at: http://localhost:8888
+
+## Notebooks
+
+The course consists of five notebooks that build upon each other:
+
+1. **01_Graph_Triplet_Extraction.ipynb**
+   - Extract structured (subject, relation, object) triplets from SEC 10-K filings
+   - Transform unstructured text into structured knowledge for a knowledge graph
+
+2. **02_LLM_Finetuning.ipynb**
+   - Fine-tune a smaller LLM (LLaMa-3 8B) for more accurate triplet prediction
+   - Leverage NVIDIA NeMo and NVIDIA Inference Microservices (NIM)
+
+3. **03_Dynamic_Database.ipynb**
+   - Set up a persistent, dynamic backend (ArangoDB) for the knowledge graph
+   - Handle triplets being added or deleted over time
+   - Create a GraphRAG agent connected to a continuously updating database
+
+4. **04_Evaluation.ipynb**
+   - Evaluate the RAG system using Ragas and NVIDIA's Nemotron-340b-reward model
+   - Assess metrics such as faithfulness, context precision, and answer relevancy
+
+5. **05_Link_Prediction.ipynb**
+   - Improve knowledge graph completeness through link prediction
+   - Use techniques like TransE model and non-negative matrix factorization
+   - Predict missing relationships within the knowledge graph
+
+## Data
+
+The notebooks use 2021 SEC documents (10-K filings) from the [Kaggle SEC EDGAR Annual Financial Filings 2021 dataset](https://www.kaggle.com/datasets/pranjalverma08/sec-edgar-annual-financial-filings-2021). The data preprocessing tools are included in the `data_prep` directory.
+
+## Requirements
+
+The key Python packages required for this project are listed in `requirements.txt` and include:
+- numpy, scipy, scikit-learn
+- torch, pykeen, fairscale
+- jupyterlab, networkx
+- ArangoDB libraries (nx_arangodb, arango)
+- NVIDIA GPU libraries (cugraph-cu12, nx_cugraph-cu12)
+- LLM frameworks (llama-index, langchain, transformers)
+- Evaluation tools (ragas, datasets)
+
+## Docker Environment
+
+The repository uses multiple containers:
+- A JupyterLab environment with CUDA support
+- NVIDIA NeMo container for model fine-tuning
+- NVIDIA NIM container for model serving
+- ArangoDB for graph database storage

community/knowledge_graph_rag/GTC25_DLI/docker-compose.yml

@@ -0,0 +1,60 @@
+version: '3'
+
+services:
+  nginx:
+    image: nginx:latest
+    ports:
+      - "${DEV_NGINX_PORT}:8888"
+    volumes:
+      - ./nginx.conf:/etc/nginx/conf.d/default.conf:ro
+      - ./notebooks:/usr/share/nginx/html:ro
+    networks:
+      - gtc25_kgrag_dli
+
+  jupyter:
+    build:
+      context: .
+      dockerfile: Dockerfile
+    volumes:
+      - ./notebooks:/workspace/notebooks
+      - ./entrypoint.sh:/entrypoint.sh
+    environment:
+      - JUPYTER_TOKEN=nvidia
+    entrypoint: /entrypoint.sh
+    networks:
+      - gtc25_kgrag_dli
+    deploy:
+      resources:
+        reservations:
+          devices:
+            - driver: nvidia
+              count: 1
+              capabilities: [gpu]
+
+  nim:
+    image: nvcr.io/nvidia/nemo-inference-microservice:23.12
+    ports:
+      - "8000:8000"
+    environment:
+      - NGC_API_KEY=$NGC_API_KEY
+      - NIM_PEFT_SOURCE=/home/nvs/loras
+      - NIM_PEFT_REFRESH_INTERVAL=3600
+      - TRANSFORMERS_CACHE=#
+      - CUDA_VISIBLE_DEVICES=2
+    volumes:
+      - ${PWD}/model/nim/:/opt/nim/.cache:rw
+      - ${PWD}/model/loras/:/home/nvs/loras:rw
+    networks:
+      - gtc25_kgrag_dli
+    deploy:
+      resources:
+        reservations:
+          devices:
+            - driver: nvidia
+              count: 1
+              capabilities: [gpu]
+
+networks:
+  gtc25_kgrag_dli:
+    name: ${COMPOSE_PROJECT_NAME}
+

community/knowledge_graph_rag/GTC25_DLI/entrypoint.sh

@@ -0,0 +1,18 @@
+#!/bin/bash
+
+# JUPYTER_TOKEN will empty in development, but set for
+# deployments. It will be applied from the environment
+# automatically when running `docker-compose up`.
+# For more details see `docker-compose.production.yml`,
+# `docker-compose.override.yml`, and
+# `nginx.conf`.
+
+jupyter lab \
+        --ip 0.0.0.0                               `# Run on localhost` \
+        --allow-root                               `# Enable the use of sudo commands in the notebook` \
+        --no-browser                               `# Do not launch a browser by default` \
+        --NotebookApp.base_url="/lab"              `# Allow value to be passed in for production` \
+        --NotebookApp.token="$JUPYTER_TOKEN"       `# Do not require token to access notebook` \
+        --NotebookApp.password=""                  `# Do not require password to run jupyter server` \
+	--notebook-dir='/workspace'
+

community/knowledge_graph_rag/GTC25_DLI/nginx.conf

@@ -0,0 +1,52 @@
+worker_processes auto;
+pid /etc/nginx/.nginx.pid;
+
+events {
+	worker_connections 768;
+}
+
+http {
+	sendfile on;
+	tcp_nopush on;
+	tcp_nodelay on;
+	client_max_body_size 0;
+	keepalive_timeout 65;
+	types_hash_max_size 2048;
+
+	default_type application/octet-stream;
+
+	ssl_protocols TLSv1 TLSv1.1 TLSv1.2; # Dropping SSLv3, ref: POODLE
+		ssl_prefer_server_ciphers on;
+
+	access_log /var/log/access.log;
+	error_log /var/log/error.log;
+
+	gzip on;
+	gzip_disable "msie6";
+
+	server {
+
+		listen 80 default_server;
+		listen [::]:80 default_server;
+
+		location / {
+			proxy_pass http://localhost/lab;
+			proxy_http_version 1.1;
+			proxy_set_header Upgrade $http_upgrade;
+			proxy_set_header Connection "Upgrade";
+			proxy_set_header Host $http_host;
+			proxy_set_header X-Forwarded-For $proxy_add_x_forwarded_for;
+			proxy_buffering off;
+		}
+
+		location /lab {
+			proxy_pass http://lab:8888;
+			proxy_http_version 1.1;
+			proxy_set_header Upgrade $http_upgrade;
+			proxy_set_header Connection "Upgrade";
+			proxy_set_header Host $http_host;
+			proxy_set_header X-Forwarded-For $proxy_add_x_forwarded_for;
+			proxy_buffering off;
+		}
+	}
+}