app.py

import gradio as gr
from huggingface_hub import InferenceClient
from transformers import BertTokenizer, BertForSequenceClassification, TrainingArguments, Trainer
import openai
import os
import faiss
import numpy as np
import requests
from datasets import load_dataset

ds = load_dataset("epfl-llm/guidelines")

# Load OpenAI and Serper API keys from Hugging Face secrets
openai.api_key = os.getenv("OPENAI_API_KEY")  # Ensure the OpenAI API key is pulled correctly
serper_api_key = os.getenv("SERPER_API_KEY")  # Ensure the Serper API key is pulled correctly

# Load PubMedBERT tokenizer and model for FDA-related processing
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 (embedding-based memory)
dimension = 768  # PubMedBERT embedding size
index = faiss.IndexFlatL2(dimension)

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)  # Ensure hidden states are returned
    hidden_state = outputs.hidden_states[-1]  # Get the last hidden state
    return hidden_state.mean(dim=1).detach().numpy()  # Take the mean across the sequence

# Example: Embed past conversation and store in FAISS
past_conversation = "FDA approval for companion diagnostics requires careful documentation."
past_embedding = embed_text(past_conversation)
index.add(past_embedding)

# Embed the incoming query and search for related memory
def search_memory(query):
    query_embedding = embed_text(query)
    D, I = index.search(query_embedding, k=1)  # Retrieve most similar past conversation
    return I

# Function to handle FDA-specific queries with PubMedBERT
def handle_fda_query(query):
    inputs = tokenizer(query, return_tensors="pt", padding="max_length", truncation=True)
    outputs = model(**inputs)
    logits = outputs.logits
    response = "Processed FDA-related query via PubMedBERT"
    return response

# Function to handle general queries using GPT-4o
def handle_openai_query(prompt):
    response = openai.Completion.create(
        engine="gpt-4o",  # Using GPT-4o as per instruction
        prompt=prompt,
        max_tokens=100
    )
    return response.choices[0].text.strip()

# Web search with Serper API
def web_search(query):
    url = f"https://google.serper.dev/search"
    headers = {
        "X-API-KEY": serper_api_key
    }
    params = {
        "q": query
    }
    response = requests.get(url, headers=headers, params=params)
    return response.json()

# Main assistant function that delegates to either OpenAI, PubMedBERT, or Serper (web search)
def respond(
    message,
    history: list[tuple[str, str]],
    system_message,
    max_tokens,
    temperature,
    top_p,
):
    # Prepare the context for OpenAI and PubMedBERT
    messages = [{"role": "system", "content": system_message}]

    for val in history:
        if val[0]:
            messages.append({"role": "user", "content": val[0]})
        if val[1]:
            messages.append({"role": "assistant", "content": val[1]})

    messages.append({"role": "user", "content": message})

    # Check if the query is related to FDA
    openai_response = handle_openai_query(f"Is this query FDA-related: {message}")
    
    if "FDA" in openai_response or "regulatory" in openai_response:
        # Search past conversations/memory using FAISS
        memory_index = search_memory(message)
        if memory_index:
            return f"Found relevant past memory: {past_conversation}"  # Return past context from memory

        # If no memory match, proceed with PubMedBERT
        return handle_fda_query(message)

    # If query asks for a web search, perform web search
    if "search the web" in message.lower():
        return web_search(message)

    # General conversational handling with GPT-4o
    response = ""
    for message in client.chat_completion(
        messages,
        max_tokens=max_tokens,
        stream=True,
        temperature=temperature,
        top_p=top_p,
    ):
        token = message.choices[0].delta.content

        response += token
        yield response


# Create Gradio ChatInterface for interaction
demo = gr.ChatInterface(
    respond,
    additional_inputs=[
        gr.Textbox(value="You are Ferris2.0, an FDA expert.", 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)")
    ],
)

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