Harness the power of the GPT Oompa Loompas to assist you effortlessly.
Yet another GPT powered agent to perform boring programming tasks.
First test version was working on less than 90 lines of code. Any code added later is just bells and whistles. You can find this first version in commit: 3f6886e
When 0613
version of GPT-3 was released, I thought that the new
function_call
functionality could be used to improve many automation tasks
such as code generation. Also olgpt would be an excuse to play with the idea of
autonomous agents building and testing code.
Previous to GPT-3 0613
version, The way everyone was using GPT-3 to generate
code was to ask for code in a "specific" output format, generate an
intermediate file containing the files and dependencies which the project will
be composed of, then asking to generate each file one by one. This approach
produces many calls to the api, each new call must contain context from previous
calls causing a waste of tokens, Also the chances of getting a bad formatted
output increases with each api call, and the missing context could also produce
files with useless functions or missing dependencies. Using 0613
version` and
some clever tricks we can generate the hole code in just one call.
flowchart TD
subgraph GPT Engineer - default
BA([Start]) --> BB[Generate\n program specs]
BB --> BC[Generate\n Unit tests]
BC --> BD[Generate\n Code]
BD --> BE[Generate\n Entry point]
BE --> BF([end])
end
subgraph smol developer
AA([Start]) --> AB[Create program structure\nlist of files]
AB --> AC[Generate\n Shared dependencies]
AC --> AD[Generate code\n for given file]
AD --> AE{Any files remaining?}
AE --> |Yes| AD
AE --> |No| AF([end])
end
subgraph olgpt
A([Start]) --> B[Generate all code\n Using function_call]
B --> E([end])
end
asdasd
flowchart LR
A([START]) --> B[Generate code]
B --> J{Improve code?}
J -->|Yes| B
J -->|No| C{Compiles\nruns?}
C -->|No| G[Extra user\n feedback]
C -->|Yes| D{Exec output\nas expected?}
D --> |Yes| E([END])
D --> |No| G
G --> F
F[Fix code] --> C