Am I wrong or can you feed it your github repo and get it to create code relevant to your project - instead of manually feeding it contextual info with every prompt?
This is a good question, because the context is limited and the more you use, the more you pay.
One technique is to use embeddings, and then only provide a subset that is close to the prompt embedding. The hard part here is chunking and deciding what close means. One example, when including a function in the context, do I need to include the body, or is the doc string and signature sufficient?
Comments
Am I wrong or can you feed it your github repo and get it to create code relevant to your project - instead of manually feeding it contextual info with every prompt?
You may be interested in this: https://github.com/mpoon/gpt-repository-loader
How does it decide which files?
This is a good question, because the context is limited and the more you use, the more you pay.
One technique is to use embeddings, and then only provide a subset that is close to the prompt embedding. The hard part here is chunking and deciding what close means. One example, when including a function in the context, do I need to include the body, or is the doc string and signature sufficient?