search

AISE-TUDelft / coco

AI Code Completions for Jetbrains and VSCode
0 stars 0 forks source link

Proivde an interface to generate upon given request #6

Closed RebelOfDeath closed 3 months ago

RebelOfDeath commented 3 months ago

Provide an interface to the application layer to be able to invoke a method of sort that generates a completion for a given request and returns it in the intended format to the application.

Ar4l commented 3 months ago

As I am extending LangChain's BaseLLM with better vllm integration, you may use the BaseLLM methods (invoke/ainvoke in this case). I will soon add the respective input/output parsers to ensure the format follows model specifications.

Ar4l commented 3 months ago

Done. We generate completions using a LangChain Runnable, so this is the interface if you must use one. In practice, this corresponds to the chain defined in server.completions.__init__.

The chain is set up in the FastAPI lifespan, which is passed to the app. This means that all models will be loaded before endpoints are set up.