app.py

import gradio as gr
import openai
import os
import requests
from transformers import BertTokenizer, BertForSequenceClassification
import torch
import faiss
import numpy as np
import json

def clean_payload(payload):
    # Remove "data:" prefix and clean newline characters
    cleaned_payload = payload.lstrip("data:").rstrip("\n")
    try:
        json_payload = json.loads(cleaned_payload)
    except json.JSONDecodeError as e:
        print(f"JSON decoding error: {e}")
        json_payload = None
    return json_payload


# API Keys and Org ID
openai.api_key = os.getenv("OPENAI_API_KEY")
openai.organization = os.getenv("OPENAI_ORG_ID")
serper_api_key = os.getenv("SERPER_API_KEY")

# Load PubMedBERT tokenizer and model
tokenizer = BertTokenizer.from_pretrained("microsoft/BiomedNLP-PubMedBERT-base-uncased-abstract")
model = BertForSequenceClassification.from_pretrained("microsoft/BiomedNLP-PubMedBERT-base-uncased-abstract", num_labels=2)

# FAISS setup for vector search
dimension = 768
index = faiss.IndexFlatL2(dimension)

# Function to embed text (PubMedBERT)
def embed_text(text):
    inputs = tokenizer(text, return_tensors="pt", truncation=True, padding="max_length", max_length=512)
    outputs = model(**inputs, output_hidden_states=True)
    hidden_state = outputs.hidden_states[-1]
    return hidden_state.mean(dim=1).detach().numpy()

# Function to retrieve info from PubMedBERT
def handle_fda_query(query):
    inputs = tokenizer(query, return_tensors="pt", padding="max_length", truncation=True, max_length=512)
    outputs = model(**inputs)
    logits = outputs.logits
    prediction = torch.argmax(logits, dim=1).item()

    # Simulate a meaningful FDA-related response
    if prediction == 1:
        return f"FDA Query Processed: '{query}' contains important regulatory information."
    else:
        return f"FDA Query Processed: '{query}' seems to be general and not regulatory-heavy."

# Function to enhance info via GPT-4o-mini
def enhance_with_gpt4o(fda_response):
    try:
        response = openai.ChatCompletion.create(
            model="gpt-4o-mini",  # Correct model
            messages=[{"role": "system", "content": "You are an expert FDA assistant."}, {"role": "user", "content": f"Enhance this FDA info: {fda_response}"}],
            max_tokens=150
        )
        return response['choices'][0]['message']['content']
    except Exception as e:
        return f"Error: {str(e)}"

def respond(message, system_message, max_tokens, temperature, top_p):
    try:
        # First retrieve info via PubMedBERT
        fda_response = handle_fda_query(message)
        
        # Stream the enhanced response via GPT-4o-mini using the client
        enhanced_response = ""
        for chat_message in client.chat_completion(...):
            payload = json.loads(chat_message.lstrip("data:").rstrip("\n"))
            enhanced_response += payload["content"]  # Or however the payload structure works

        # Return both the PubMedBERT result and the enhanced version
        return f"Original Info from PubMedBERT: {fda_response}\n\nEnhanced Info via GPT-4o-mini: {enhanced_response}"

    except Exception as e:
        return f"Error: {str(e)}"

# Gradio Interface
demo = gr.Interface(
    fn=respond,
    inputs=[
        gr.Textbox(label="Enter your FDA query", placeholder="Ask Ferris2.0 anything FDA-related."),
        gr.Textbox(value="You are Ferris2.0, the most advanced FDA Regulatory Assistant.", label="System message"),  
        gr.Slider(minimum=1, maximum=2048, value=512, step=1, label="Max new tokens"),
        gr.Slider(minimum=0.1, maximum=4.0, value=0.7, step=0.1, label="Temperature"),
        gr.Slider(minimum=0.1, maximum=1.0, value=0.95, step=0.05, label="Top-p (nucleus sampling)")
    ],
    outputs="text",
)

if __name__ == "__main__":
    demo.launch()