07-12, 11:50–12:00 (Europe/Amsterdam), REPL (2, main stage)
How can you teach ChatGPT to write well informed Julia code? In this talk we’ll address the basic retrieval augmented generation (RAG) ideas that back JuliaHub’s AskAI feature; one of the most popular code generation tools for Julia.
JuliaHub’s AskAI feature has been generating good Julia code since the Spring of 2023. To power it up, we are processing a big archive of julia knowledge, and we are combing through it, to find the most relevant knowledge regarding the user’s query. In this talk we will discuss issues with Julia documentation maintenance, the splitting of this knowledge into chunks the retrieval and the context engineering required to achieve improved performance over plain ChatGPT. We’ll show that our methods improve ChatGPT (3.5) at least over two dimensions:
- serving the latest relevant information and
- reducing the hallucinations that make ChatGPT think that Julia is just python.jl
Finally we’ll discuss limitations and next steps in Julia code generation, along with potential integrations into IDEs (Pluto, Jupyter)
Misc at Pluto.jl, Software engineering JuliaHub (Ask AI, Pluto integrations, Julia Web applications & JavaScript). In the past, I've been a software engineer, a business analyst, a consultant, a data entry intern, a waiter and a sailor. Very concerned about climate change, trying to do something about it (algae bioreactors in the making).