EduQuery

Local Setup

1. Install Requirements using pip install -r requirements.txt
2. Setup Environment Variables: File dummy.env has the template for keys. Rename the file to .env after adding the api keys and let it be in the same root folder.
   • You have to set up API keys for Neo4j and Gemini LLM API
   • ./config/app_config.yaml has the app settings like choice of LLM
     • If you want use OpenAI models instead, set use_llm: openai in this file
     • OPTIONAL:
     • If you want to use a different LLM:
       1. Add the config like other LLMs in ./config/app_config.yaml
       2. Setup API keys for the new LLM in .env
       3. And LLM initialization code in ./src/llm/llm_core.py: LLMFactory class
3. Setup Database:
   1. If you are setting up the Neo4j Database for the first time, you need to install the
   2. Run the script python scripts/setup_neo4j.py to populate the Knowledge Graph DB

The demo has both CLI mode

• For the CLI mode: run python demo.py use --verbose for tracing the pipeline

Knowledge Graph Schema

Why Graph DB over Relational DB?

• This was a design decision based on the following points:
• As the use-cases grow, capturing more complicated relationships will be easier in KG
  • Example: Group Projects, Prerequisites, TAs etc.
• Graph DBs provides flexible Schema
• Graph DBs can traverse relationships faster than SQL joins.

How does the pipeline work?

Directions for Expansion

1. Adding more courses
   • This is already supported. Just put more course folders in data folder and run python scripts/setup_neo4j.py
   • This will skip any course folders that are already processed.
   • Will not duplicate actors nodes like Student nodes and Instructor nodes (assuming they all have a unique id like in any university)
2. Using More Examples for LLM
   • As the number of examples increase we can add an example sector module which dynamically loads examples from a vector store.
   • In the src/datamodel/graph_db.py: CypherQueryRepository load the examples in a vector store
   • Add the API keys for the vector store in .env file and application level configurations in config/app_config.yaml
   • Create a class to use these configuraitons in src/api_keys.py
   • Use SemanticSimilarityExampleSelector in Step 3 of the pipeline src/pipeline/edu_query.py
   • Similarly select an encoder model for encoding the example questions and setup API keys and App configs in the same way as explained
3. Use a different backend database
   • Setup DB code in src/datamodel module. Use graph_db.py for reference
   • Create a new pipeline in /src directory. The abtract class src/pipeline/edu_query: EduQuery. Follow src/graph_pipeline.py for reference.

Deployment Using Docker

This backend is containerized with Docker and can be deployed to any cloud platform that supports containerized applications, such as Google Cloud Run, AWS ECS, Azure Container Apps.

Tech Stack:

1. Neo4J
2. LangChain
3. Python
4. Google Gemini LLM