Skip to content

Latest commit

 

History

History
104 lines (76 loc) · 2.72 KB

VGPU_SIMPLIFIED_SETUP.md

File metadata and controls

104 lines (76 loc) · 2.72 KB

Simplified vGPU RAG System

Overview

The vGPU RAG system has been simplified to provide a seamless experience:

  • NO collection selection - just one pre-loaded knowledge base
  • NO manual uploads - all PDFs are automatically ingested
  • NO configuration - everything works out of the box

How It Works

  1. Single Knowledge Base: All vGPU documentation is loaded into one collection called vgpu_knowledge_base
  2. Automatic Loading: When the system starts, it automatically ingests ALL PDFs from the vgpu_docs folder
  3. Enhanced Validation: The system still validates vGPU profiles and provides accurate recommendations

Setup Instructions

1. Add Your PDFs

Place ALL your NVIDIA vGPU documentation PDFs in the vgpu_docs folder:

vgpu_docs/
├── nvidia_vgpu_software_user_guide.pdf
├── vgpu_profile_specifications.pdf
├── a40_datasheet.pdf
├── l40s_specifications.pdf
├── esxi_vgpu_deployment_guide.pdf
└── [any other vGPU PDFs]

2. Start the System

# Set your NGC API key
export NGC_API_KEY="nvapi-your-key-here"

# Start everything (cloud mode)
./scripts/start_vgpu_rag.sh --skip-nims

3. Use the System

Open http://localhost:8090 and start asking questions!

Examples:

  • "What vGPU profiles are available for A40 GPUs?"
  • "I have 4x L40S GPUs, how many VMs can I run?"
  • "Compare vGPU vs passthrough for AI workloads"

What Changed?

Before (Complex)

  • Multiple collections to manage
  • Manual collection selection in UI
  • Complex document organization
  • Users had to know which collection to search

After (Simple)

  • One collection with everything
  • No UI clutter
  • Drop PDFs in one folder
  • System automatically finds relevant content

Benefits

  1. Easier Setup: Just drop PDFs and go
  2. Better User Experience: No confusion about collections
  3. Same Intelligence: Still validates profiles and provides enhanced recommendations
  4. Faster Onboarding: New users can start immediately

Technical Details

  • Collection Name: vgpu_knowledge_base
  • Bootstrap automatically creates collection and ingests all PDFs
  • Enhanced validation still works (profile checking, capacity calculations)
  • All PDFs are searched for every query

Troubleshooting

PDFs Not Loading?

# Check if PDFs exist
ls -la vgpu_docs/*.pdf

# Re-run bootstrap
docker compose -f deploy/compose/docker-compose-bootstrap.yaml up

Want to Update PDFs?

  1. Add/remove PDFs in vgpu_docs
  2. Re-run bootstrap to update the knowledge base

Need to Reset?

# Stop everything
./scripts/stop_vgpu_rag.sh

# Remove volumes (careful - this deletes all data!)
docker volume rm nvidia-rag_milvus-data

# Start fresh
./scripts/start_vgpu_rag.sh --skip-nims