Set Up for my own libraries

Set up virtual environment with virtualenv venv and activate it source venv/bin/activate a. Note: Windows commands are slightly different
In the virtual environment shell, run pip install -r requirements.txt
To add newly installed libraries to requirements.txt, can fun pip freeze > requirements.txt after installation

Server

To run the server: uvicorn main:app --reload a. Runs on port 8000 by default but can specify with uvicorn main:app --reload --port 8000 b. http://localhost:8000/docs contains openapi (swagger) c. http://localhost:8000/redoc contains Redocs d. http://localhost:8000/openapi.json contains the openapi specification automatically created for you by fastapi based on your routes

Notes

Requirements.txt has too many libraries for production but is fine for now. Clean up unnecessary ones

Deployment

Run gcloud app deploy
To view service, run gcloud app browse
Ran gcloud auth application-default login to authenticate google secret manager

Google Notes

Agent Builder allows you to easily add additional data to an LLM
- Can build an LLM with access to your FHIR store on Agent Builder
- Doesn't look like apps built with Agent Builder can be deployed via API
Google AI Studio is a GUI to help with prompt engineering. Ex: can give it expected inputs and outputs and then it feeds those as examples into the model.
VertexAI

Summary

Fine-Tuning: modify model weights based on a labeled data set
RAG: provide additional information to the model before it replies. Need to search for the correct data
- On Google, you can use their no-code tools to give it data, and then it does a semantic search for you
- You'd need to do more of this on your own if you weren't using google's infrastructure