Skip to content

demo-script.md

Latest commit

 

History

History
91 lines (59 loc) · 2.4 KB

demo-script.md

File metadata and controls

91 lines (59 loc) · 2.4 KB

10-Minute Demo Script

1. Introduction

Say that this is a full-stack multi-modal Graph RAG system that supports:

  • text
  • PDFs
  • images

and combines:

  • vector retrieval using Pinecone
  • graph reasoning using NetworkX
  • local Ollama generation for grounded answers

2. Architecture Overview

Show the Mermaid diagram from the README and explain the pipeline in one line each:

  • ingestion API receives files
  • modality-specific extraction converts them to retrieval-ready text
  • chunks are stored in Pinecone and images are embedded directly with CLIP
  • entities and relationships are stored in the knowledge graph
  • user query triggers retrieval + graph expansion + answer generation

3. Ingestion Demo

Upload:

  • one .txt file
  • one .pdf
  • one .jpeg or .png

While doing that, say:

  • text and PDF are parsed directly
  • images are directly embedded using a CLIP visual model
  • text descriptions are embedded using the same CLIP model, putting everything in the same search space

Then highlight in the UI:

  • inventory cards
  • file counts
  • chunk counts
  • entity counts
  • graph node and edge totals

4. Query Demo

Ask a cross-modal question such as:

Find me running shoes that look like this uploaded image, and summarize their reviews.

Then point out:

  • the answer panel
  • citations
  • graph insights
  • retrieved evidence cards

Mention that the answer is grounded in retrieved context instead of direct free-form generation.

5. Inventory Management Demo

Remove one uploaded file from the inventory using the Remove File button.

Then show:

  • the card disappears
  • the graph metrics update
  • the removed file no longer affects later answers

This helps demonstrate lifecycle management, not just ingestion.

6. Engineering Choices

Briefly explain:

  • Pinecone was chosen for serverless cloud vector storage to save local RAM
  • NetworkX was chosen for lightweight graph reasoning without extra operational overhead
  • qwen2:0.5b was chosen for lightweight local answer generation
  • sentence-transformers/clip-ViT-B-32 was chosen to demonstrate true cross-modal vector embedding
  • local Ollama was used to avoid cloud quota issues during live demo

7. Close

Finish by connecting the design to the literature survey:

  • modern agentic systems rely on memory, planning, and tool orchestration
  • this project applies that idea through vector retrieval, graph memory, and orchestrated grounded generation