CodeQA

Problem Statement:
Given a URL of a Github repository, the proposed solution enables the user to have a priliminary understanding of the repository by asking the system in a conversation style setup.

High Level Design:

3 main modules, each of which function as a standlone dockerized microservice:

• API: That captures the main tasks of downloading the repository, processing it, embedding the same and so on. More details can be found here
• UI: Simple interface to see the solution in action.
• VecDB: A vector database that supports CRUD of vector embeddings as well as some metadata information.

Demo:

See demo here

QuickStart:

• Include a .env file with a key for OPENAI_API_KEY="" in the API module (ie in this path)
• Download the embedding model from here and place it in this location
• The microservices are encapsulated as composable docker services. Hence run
  docker-compose up ---build
  at the root location.
• You can find each of the modules in the following URLs:
  • UI: localhost:8000
  • API: localhost:8001/docs
  • Qdrant Dashboard: localhost:6333/dashboard

More technical details:

• UI: Streamlit is used for building a basic interactive app. Screenshot:
  
  
• API: Fast API is used for implementing RESTful services. More details about the supported APIs can be found here.
• VectorDB: Self-hosted Qdrant database is used as a vector database.

Next steps:

• Improvements can be done at several places, Some of them (but not limited to) could be:
  • Filtering strategies while downloading the repo
  • Detecting programming language of the scripts and performing appropriate cleaning strategies
  • Creating richer metadata for the scripts at both document and chunk level such as summaries of functions, comments, function names etc.
  • More advanced strategies while embedding the scripts, to bring about "Contextual RAG".
• I have tried to add comments (@TODO) in appropriate places in the scripts.