app.py

import os
import torch
import gradio as gr
from transformers import AutoTokenizer, AutoModelForCausalLM

# =====================================================
# Environment setup
# =====================================================
os.environ["TRANSFORMERS_CACHE"] = "/tmp/hf_cache"
os.environ["HF_HOME"] = "/tmp/hf_home"

# =====================================================
# Model configuration
# =====================================================
GEN_MODEL = "hackergeek/qwen3-harrison-rag"
HF_TOKEN = os.getenv("HF_TOKEN")

if not HF_TOKEN:
    print("⚠️ No Hugging Face token found. Set one using:")
    print("   export HF_TOKEN='your_hf_token_here'")

# =====================================================
# Load private model
# =====================================================
def load_private_model(model_name, token):
    dtype_value = torch.float16 if torch.cuda.is_available() else torch.float32
    load_kwargs = {
        "dtype": dtype_value,
        "cache_dir": "/tmp/hf_cache",
        "low_cpu_mem_usage": True,
    }
    try:
        import accelerate
        load_kwargs["device_map"] = "auto"
    except ImportError:
        print("⚠️ `accelerate` not installed — default device placement used.")

    tokenizer = AutoTokenizer.from_pretrained(model_name, token=token)
    model = AutoModelForCausalLM.from_pretrained(model_name, token=token, **load_kwargs)
    return tokenizer, model

tokenizer, model = load_private_model(GEN_MODEL, token=HF_TOKEN)

# =====================================================
# Dynamic token allocation
# =====================================================
def calculate_max_tokens(query, min_tokens=1000, max_tokens=8192, factor=8):
    query_tokens = len(tokenizer(query)["input_ids"])
    dynamic_tokens = query_tokens * factor
    return min(max(dynamic_tokens, min_tokens), max_tokens)

# =====================================================
# Generate long, complete, structured answers
# =====================================================
def generate_answer(query, history):
    if not query.strip():
        return history, history

    # Correct common typos
    corrected_query = query.replace("COPP", "COPD")

    # Step 1: Rephrase for precise retrieval
    rephrase_prompt = (
        "You are a medical assistant. Rephrase this query for precise retrieval:\n\n"
        f"Query: {corrected_query}\n\nRephrased query:"
    )
    inputs = tokenizer(rephrase_prompt, return_tensors="pt").to(model.device)
    rephrased_ids = model.generate(**inputs, max_new_tokens=128, do_sample=False)
    rephrased_query = tokenizer.decode(
        rephrased_ids[0], skip_special_tokens=True
    ).split("Rephrased query:")[-1].strip()

    # Step 2: Generate detailed structured answer
    max_tokens = calculate_max_tokens(rephrased_query)
    prompt = (
        "You are a retrieval-augmented medical assistant. Provide a **long, detailed, structured** medical answer "
        "as if writing a concise clinical guideline. Use markdown headings and bullet points. "
        "Each section should include multiple complete sentences and clear explanations.\n\n"
        "Follow this structure:\n"
        "### Definition / Description\n"
        "### Epidemiology / Causes\n"
        "### Symptoms & Signs\n"
        "### Diagnosis / Investigations\n"
        "### Complications\n"
        "### Treatment & Management\n"
        "### Prognosis / Prevention\n"
        "### Key Notes / References\n\n"
        "At the end, include a **🩺 Quick Summary** with 3–5 key takeaways written in plain English "
        "that a non-medical reader could understand.\n\n"
        f"User query: {rephrased_query}\n\nAnswer:"
    )

    inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
    output_ids = model.generate(
        **inputs,
        max_new_tokens=max_tokens,
        do_sample=True,
        temperature=0.8,
        top_p=0.9,
        repetition_penalty=1.2,
        pad_token_id=tokenizer.eos_token_id,
    )

    output = tokenizer.decode(output_ids[0], skip_special_tokens=True)
    answer = output.split("Answer:")[-1].strip()

    # Clean up potential triple breaks
    while "\n\n\n" in answer:
        answer = answer.replace("\n\n\n", "\n\n")

    history = history + [(query, answer)]
    return history, history

# =====================================================
# Gradio interface
# =====================================================
with gr.Blocks(title="Qwen3-Harrison-RAG Chatbot") as demo:
    gr.Markdown("""
    # 🧠 Qwen3-Harrison-RAG Medical Chatbot  
    This model provides **guideline-style medical answers** with structured sections and a **Quick Summary**.  
    *For educational and informational purposes only — not a substitute for professional medical advice.*
    """)
    chatbot = gr.Chatbot(height=480, show_label=False)
    with gr.Row():
        msg = gr.Textbox(placeholder="Ask a detailed medical question...", scale=4)
        clear = gr.Button("Clear", scale=1)
    msg.submit(generate_answer, [msg, chatbot], [chatbot, chatbot])
    clear.click(lambda: None, None, chatbot, queue=False)

# =====================================================
# Launch
# =====================================================
if __name__ == "__main__":
    demo.launch(server_name="0.0.0.0", server_port=int(os.environ.get("PORT", 7860)), debug=True)