This project provides a Streamlit-powered web interface to parse PDF files and generate structured clinical summaries using OpenAI models (via LiteLLM). It includes end-to-end evaluation with ROUGE and hallucination-free scorers, plus an interactive feedback system (thumbs up/down and notes) logged through Weights & Biases Weave.

Create a .env file with your API keys:

OPENAI_API_KEY=your_openai_api_key
ANTHROPIC_API_KEY=your_anthropic_api_key
WANDB_API_KEY=your_wandb_api_key
WANDB_ENTITY=your_wandb_team_name

# OPTIONAL: Set WANDB_BASE_URL if you are writing to a non wandb.ai deployment
WANDB_BASE_URL=your_wandb_host

Install dependencies:

pip install -r requirements.txt

Contains core model classes and utility functions for PDF parsing, LLM chat interaction, and summarization logic. Defines the ChatModel, AuthoringModel, and related Weave Ops used by both the Streamlit app and evaluation script.

How to use:
This file is imported by streamlit.py and evaluation.py and is not meant to be run directly.

A Streamlit web application for uploading PDF files, generating clinical summaries using LLMs, and collecting user feedback. The app parses PDFs, summarizes their content with a configurable system prompt, and displays results with interactive feedback buttons.

How to run:

streamlit run streamlit.py

A script for automated evaluation of different LLM summarization models on a clinical dataset. It uses Weave's evaluation framework, computes ROUGE-L scores, and leverages hallucination and summarization scorers. Results are logged for comparison across models.

How to run:

python evaluation.py

• Ensure you have the proper API keys set up if you're evaluating both OpenAI and Anthropic models.
• All scripts require dependencies listed in requirements.txt and are best run inside the provided conda environment.