Detailed description:
Today, Large Language Models (LLMs) show fascinating abilities when it comes to generating code. However, they have a major shortcoming: they learn coding by sheer repetition and develop their zero-shot capabilities through studying example code. This prevents the model from scaling beyond what is included in the training data as well as using versions of packages/libraries that emerged after the model was trained.
Hence, we propose a Retriever Augmented Generation (RAG) framework that accesses a vector database with the encoded documentation of every function of a package. On request, the agent will match the user input query to the function that is most likely needed in order to generate the code for achieving what the user wants to do. This enables up-to-date knowledge of e.g. Python libraries like Pandas without expensive retraining of the LLM.
Group members: Jan Henrik Bertrand, André Gilbert, Felix Noll, Marc Grün
Detailed description: Today, Large Language Models (LLMs) show fascinating abilities when it comes to generating code. However, they have a major shortcoming: they learn coding by sheer repetition and develop their zero-shot capabilities through studying example code. This prevents the model from scaling beyond what is included in the training data as well as using versions of packages/libraries that emerged after the model was trained. Hence, we propose a Retriever Augmented Generation (RAG) framework that accesses a vector database with the encoded documentation of every function of a package. On request, the agent will match the user input query to the function that is most likely needed in order to generate the code for achieving what the user wants to do. This enables up-to-date knowledge of e.g. Python libraries like Pandas without expensive retraining of the LLM.
Group members: Jan Henrik Bertrand, André Gilbert, Felix Noll, Marc Grün
GitHub Link: https://github.com/Andre-Gilbert/llms-using-rag-for-docs