app.py

import os
import gradio as gr
from transformers import pipeline
from elasticsearch import Elasticsearch

# Connect to Elasticsearch
es = Elasticsearch(hosts=["https://data.neuralnoise.com:9200"],
                   basic_auth=('elastic', os.environ['ES_PASSWORD']),
                   verify_certs=False, ssl_show_warn=False)

def search(query, index="pubmed", num_docs=3):
    """
    Search the Elasticsearch index for the most relevant documents.
    """
    
    print(f'Running query: {query}')
    
    es_request_body = {
        "query": {
            "match": {
                "content": query  # Assuming documents have a 'content' field
            }
        }, "size": num_docs
    }

    response = es.options(request_timeout=60).search(index=index, body=es_request_body)
    
    # Extract and return the documents
    docs = [hit["_source"]["content"] for hit in response['hits']['hits']]

    print(f'Received {len(docs)} documents from index {index}')

    return docs

def rag_pipeline(prompt, index="pubmed", num_docs=3, model_name="HuggingFaceH4/zephyr-7b-beta"):
    """
    A simple RAG pipeline that retrieves documents and uses them to enrich the context for the LLM.
    """
    # Load your language model from HuggingFace Transformers
    generator = pipeline("text-generation", model=model_name)

    num_docs = int(num_docs)

    # Retrieve documents
    docs = search(prompt, index=index, num_docs=num_docs)
    joined_docs = '\n\n'.join(docs)
    
    messages = [
        {
            "role": "system",
            "content": f"You are an advanced medical support assistant, designed to help clinicians by providing quick access to medical information, guidelines, and evidence-based recommendations. Alongside your built-in knowledge, you have access to a curated set of documents retrieved from trustworthy sources such as Wikipedia and PubMed. These documents include up-to-date medical guidelines, research summaries, and clinical practice information. You should use these documents as a primary source of information to ensure your responses are based on the most current and credible evidence available. Your responses should be accurate, concise, and in full compliance with medical ethics. You must always remind users that your guidance does not substitute for professional medical advice, diagnosis, or treatment. Your tone should be professional, supportive, and respectful, recognizing the complexity of healthcare decisions and the importance of personalized patient care. While you can offer information and suggestions based on the documents provided and current medical knowledge, you must emphasize the importance of clinicians' expertise and judgment in making clinical decisions.\n\nRetrieved documents from {index}:\n\n{joined_docs}"
        }, {
            "role": "user",
            "content": prompt
        }
    ]

    for message in messages:
        print('MSG', message)

    # Generate response using the LLM
    response = generator(messages, max_new_tokens=256, return_full_text=False)
    
    # Return the generated text and the documents
    return response[0]['generated_text'], joined_docs

# Create the Gradio interface
iface = gr.Interface(fn=rag_pipeline,
                     inputs=[
                         gr.Textbox(label="Input Prompt", value="Are group 2 innate lymphoid cells (ILC2s) increased in chronic rhinosinusitis with nasal polyps or eosinophilia?"),
                         gr.Dropdown(label="Index", choices=["pubmed", "wikipedia", "textbooks"], value="pubmed"),
                         gr.Number(label="Number of Documents", value=3, step=1, min_value=1, max_value=10),
                         gr.Dropdown(label="Model", choices=["HuggingFaceH4/zephyr-7b-beta", "meta-llama/Llama-2-7b-chat-hf", "meta-llama/Llama-2-13b-chat-hf", "meta-llama/Llama-2-70b-chat-hf"], value="HuggingFaceH4/zephyr-7b-beta")
                     ],
                     outputs=[
                         gr.Textbox(label="Generated Text"),
                         gr.Textbox(label="Retrieved Documents")
                     ],
                     description="Retrieval-Augmented Generation Pipeline")

# Launch the interface
iface.launch()