Finetuning with the whole GitLab codebase and updating it
Code explanations from the Duo Chat are quite shallow. I still have to navigate files manually and figure out where symbols are. Sourcegraph doesn't help much too, because Ruby codebase is hard to parse algorithmically. Large neural network model seems to be the ideal choice to load 1 Gb structured data of GitLab sources. The questions are:
- What would be the ideal model to upload code into?
- How to upload it (is that fine tuning)?
- How to keep the knowledge updated?