Make GPT-Pilot Smarter with Few-shot Retrieval

[MaxAnfilofyev]

GPT-Pilot often gets into a loop when trying to resolve issues.
Providing relevant examples in a prompt has shown to improve performance of GPT-4 on multiple benchmarks.

1. Identify most relevant examples of how a similar issue has been solved in the past. The simplest approach could be if-thens. A more sophisticated approach could be to use embeddings and search to return the most relevant examples.
2. When an issue that took GPT several tries has been resolved, summarize the issue and the solution and store in a vector DB so that the solution can be referenced when a similar issue arises. These solutions could be shared between GPT-Pilot users if the DB is centralized.

Reference implementation: https://www.youtube.com/watch?v=UqYzzjTmKj8