I built this tool as a fun project to learn Retrieval Augmented Generation (RAG) but it turned out to be much more complicated than expected. Humour, in particular xkcd comics, require high levels of intelligence and also relevant knowledge. RAG doesn't actually seem perfectly suitable for it because the retrieval is done using a much simpler model than the latest LLMs.
While playing around, I found it to be decent most of the time. Sometimes, it would suggest a comic which isn't perfectly suitable but has some overlap atleast. It took quite a few experiments to find the current combination of embedding model and llm as well as the retrieval algo and related params.
