app.py

import os
import torch
import re
import gradio as gr
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig, TextIteratorStreamer
from peft import PeftModel
from threading import Thread

# ==========================================
# 1. SETUP & AUTHENTICATION
# ==========================================
# Mistral requires an HF token to download the base model. 
# You must add your token in the HF Spaces "Settings" -> "Variables and secrets" as 'HF_TOKEN'
hf_token = os.environ.get("HF_TOKEN")

BASE_MODEL = "mistralai/Mistral-7B-Instruct-v0.2"
ADAPTER_REPO = "your-username/Medical-Mistral-7B-LoRA" # <--- CHANGE THIS TO YOUR REPO

print("Booting up Clinical AI Server...")

tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL, token=hf_token)
tokenizer.pad_token = tokenizer.eos_token

# Load Base Model in 4-bit precision
bnb_config = BitsAndBytesConfig(
    load_in_4bit=True, 
    bnb_4bit_quant_type="nf4", 
    bnb_4bit_compute_dtype=torch.float16
)

base_model = AutoModelForCausalLM.from_pretrained(
    BASE_MODEL, 
    quantization_config=bnb_config, 
    device_map="auto",
    token=hf_token
)

# Merge Base Model with your Trained LoRA Adapter
print("Downloading and attaching LoRA Adapter from Hub...")
model = PeftModel.from_pretrained(base_model, ADAPTER_REPO, token=hf_token)
model.eval()
print("✅ Model Ready!")

# ==========================================
# 2. FORMATTING FUNCTION
# ==========================================
def format_and_clean(text):
    end_triggers = ["regards", "hope this", "hope i", "let me know", "dr.", "thanks for"]
    lower_text = text.lower()
    cutoff = len(text)
    
    for trigger in end_triggers:
        idx = lower_text.find(trigger)
        if idx != -1 and idx < cutoff:
            cutoff = idx
            
    clean = text[:cutoff].strip().replace('*', '')
    
    sentences = clean.split(". ")
    capitalized = [s.capitalize() for s in sentences if s]
    clean = ". ".join(capitalized)

    clean = re.sub(r'(?i)assessment\s*:', '**Assessment:** ', clean)
    clean = re.sub(r'(?i)analysis\s*:', '\n\n**Analysis:** ', clean)
    clean = re.sub(r'(?i)recommended action\s*:', '\n\n**Recommended Action:** ', clean)
    
    return clean

# ==========================================
# 3. CHAT ENGINE & UI
# ==========================================
def clinical_chat(message, history):
    prompt = f"""<s>[INST] You are a highly intelligent clinical AI. The user says: "{message}"
    
    If the user provides symptoms, diagnose them. If the user asks about a known condition, provide treatment advice.
    
    You MUST format your response exactly like this:
    **Assessment:** [Name of the condition]
    **Analysis:** [Brief explanation]
    **Recommended Action:** [Medical advice]
    
    Do not include greetings, sign-offs, or links. [/INST]"""
    
    inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
    streamer = TextIteratorStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True, timeout=10.0)
    
    generation_kwargs = dict(
        **inputs,
        streamer=streamer,
        max_new_tokens=300,
        temperature=0.2, 
        repetition_penalty=1.15, 
        do_sample=True,
        pad_token_id=tokenizer.eos_token_id
    )
    
    thread = Thread(target=model.generate, kwargs=generation_kwargs)
    thread.start()
    
    partial_text = ""
    for new_token in streamer:
        partial_text += new_token
        yield format_and_clean(partial_text)

# Launch UI inside Docker container
theme = gr.themes.Soft(primary_hue="blue", neutral_hue="slate")

demo = gr.ChatInterface(
    fn=clinical_chat,
    title="⚕️ Clinical AI Diagnostic Assistant",
    description="**Enter your symptoms or medical queries below for a professional analysis.**",
    theme=theme,
    examples=[
        "I am having severe headache, body pain and strain in my neck.",
        "What medicine should I take if I have high cholesterol?",
        "I have been sneezing and coughing for 3 days."
    ]
)

if __name__ == "__main__":
    # 0.0.0.0 and port 7860 are strictly required for Docker on Hugging Face Spaces
    demo.launch(server_name="0.0.0.0", server_port=7860)