Digital China Worldwide

Exploring RAG for Resource Recommendation

Kwok-leong Tang

Harvard University

2025-03-13

The Origin of the Digital China Worldwide

  • The 2023 Tools of the Trade conference at Harvard University
  • Featured over 130 innovative projects
  • Many intriguing projects remained unknown to the Chinese Studies community
  • Prof. Peter Bol initiated to build a directory of these projects.

Another Directory Project?

Luce/ACLS China Studies Digital Mapping Project

Turn Digital China Worldwide into a Chatbot

Retrieval-Augmented Generation

flowchart TB
    subgraph Direct["Direct LLM Query"]
        direction LR
        U1[User Query] --> L1[LLM] --> R1[Response]
        style L1 fill:#ff9e2c,stroke:#39332c,color:#14110f
    end


flowchart TB

    subgraph RAG["RAG-Enhanced Query"]
        direction LR
        U2[User Query] --> RC[Retrieval Component]
        RC --> |Search| DB[(Knowledge Base)]
        DB --> |Relevant Documents| RC
        RC --> CP[Context Processor]
        CP --> |Enhanced Prompt| L2[LLM]
        L2 --> R2[Response]
        style L2 fill:#ff9e2c,stroke:#39332c,color:#14110f
        style DB fill:#97500c,stroke:#39332c,color:#f4f2ec
        style RC fill:#2a251f,stroke:#39332c,color:#ece7dd
        style CP fill:#2a251f,stroke:#39332c,color:#ece7dd
    end

    classDef default fill:#1b1714,stroke:#39332c,stroke-width:2px;

Knowledge Base

Tech Stack

  • GenAI Workflow:Flowise (Self-hosted)
  • GenAI Model: OpenAI GPT-4o-mini through Harvard endpoints
  • Resource Data Collaboration: NocoDB (Self-hosted)
  • Vector Database: Qdrant (Self-hosted)

Demo of the DCW Chatbot

https://dcw.fairbank.fas.harvard.edu

Problems

  • “Not smart enough”, “Doesn’t respond well”
  • Scraping and upsert mechanism
  • Data quality (multilingual problems)

Next Steps

  • Expand the knowledge base
  • Please suggest more resources, especially the projects from your institutions
  • Please give us feedback
  • Exploration of collaborations

Thank You Very Much!