GenAI for Databases

Digital China Initiative, Harvard University

2025-02-21

Digital China Initiative

• An initiative of Harvard University to prompt digital tools and methods for Chinese Studies.
• PIs: Prof. Peter Bol and Prof. Michael Szonyi
• Provide support to Harvard faculty, students, and researchers in Chinese Studies.

The GenAI Turn

• In 2022, the release of ChatGPT (and other Large Language Models, LLMs) has brought generative AI turn to the usage of digital tools and methods in Chinese Studies.
• In April 2023, CBDB and DCI held the first GenAI workshop for Chinese Studies and humanties at Harvard University.
• We decided that our future projects should incline to the GenAI approach.
• DCI has hosted a series of GenAI workshops for Chinese Studies since 2023.

Projects Involving GenAI

Nehru Papers

• This is a NEH project that led by Tansen Sen (NYU), Gal Gvili (McGill), and Arunabh Ghosh (Harvard).
• It is a project supported by Digital China Initiative.
• The project has to digitize documents collected from the Indian archives which are related to Sino-Indian relations during the 60s and 70s.
• In this stage, we are focusing on creating a catalog for the documents (~50K+?).

Nehru Papers

• Most of the documents are in English, a few are in Indian languages and Chinese.
• Tech stacks
  • OCR: done at NYU Shanghai
  • Data extraction: Self-hosted n8n + OpenAI models
  • Data storage and collaboration: Self-hosted Nocodb

Data Extraction Workflow

Reviewing First Draft

Collaborating

Standardization

Next Steps

• The current experiment are limited to 3000 English documents.
• We hope that we will have a few dictionaries (code tables) for entities after expert review.
• Building an API for evaluating and standardizing the extracted entities.
• Import the data into two databases: the NYU Shanghai China-India database and the Woodrow Wilson Center database.

Digital China Worldwide Chatbot

• The project was directory project that orginated at the Tools of Trade: the Way Forward conference (hosted by DCI) in 2022.
• The project seems to be overlapping with the ACLS China Studies Digital Mapping Project.
• We decided to make the project as a RAG-based chatbot. Besides our original directory, we also include the data from the China Studies Digital Mapping Project.
• You can access the first version here: https://dcw.hudci.org

Digital China Worldwide Chatbot

• Tech stacks (current version)
  • Qdrant for vector database
  • Chanlit for interface
  • OpenAI gpt-4o-mini model
  • Nocodb for data storage and collaboration
• Tech stacks (future version)
  • Qdrant for vector database
  • Flowise for chat workflow
  • n8n for api and interface (integrate with Slack and other tools)
  • Nocodb for data storage and collaboration
  • Scraping machanism

Monitoring GenAI Applications

• LiteLLM: https://litellm.ai/
  • Monitoring the usage of LLM APIs.
  • Creating API keys for different projects, people, and apps.
• Langfuse: https://langfuse.com/
  • Monitoring GenAI applications.
  • Recording tokens usage, model cost and other metrics.

Challenges

• The cost for building a GenAI apps are different from “traditional” databases and platforms. Every actions (e.g., query, update, delete) will cost tokens/money.
• Building and maintaining approaches. For example, fine-tune models vs. RAG or Agentic?
• Rapid development of all aspect of GenAI. How to build in a flexible way?