tutor.py

import os
import json
import torch
from threading import Thread
from transformers import TextIteratorStreamer

def generate_vqa_response(model, tokenizer, message, history, modality, image):
    """
    Generates an educational scenario and Socratic questions via local LLM.
    IMPLEMENTED AS GENERATOR FOR STREAMING.
    """
    if model is None or tokenizer is None:
        yield "### Error\nFailed to load local MedGemma model. Please verify your HF_TOKEN and log output."
        return
        
    if image is None:
        focus = "physiological markers, systemic interactions, and clinical diagnostic criteria"
        system_prompt = f"""You are the Clinical Generalist, a highly specialized medical tutoring AI. Your purpose is to facilitate clinical reasoning, not just provide answers. 

Follow this structural protocol for EVERY generation:
1. CLINICAL OVERVIEW: Provide a brief, concise overview of the pathophysiology or concepts surrounding the query.
2. SYSTEMIC INVENTORY: Explicitly focus on {focus}.
3. DIFFERENTIAL REASONING: Mention the primary differential diagnosis but immediately contrast it with a 'mimic'.
4. SOCRATIC QUESTIONING: Answer the clinician's query Socraticly. Challenge the clinician to justify their reasoning.

Tone: Professional, objective, and Socratic. Ensure you provide a complete answer."""

    else:
        if modality == "Chest X-Ray" or modality == "X-Ray":
            focus = "costophrenic angles, hilar shadows, and cardiac silhouette"
        elif modality == "CT Scan":
            focus = "Hounsfield Units (HU), axial cross-sections, and windowing (Lung vs. Soft Tissue)"
        elif modality == "MRI":
            focus = "T1/T2 weighted signals, contrast enhancement, and multi-planar viewing"
        else:
            focus = "key anatomical landmarks"

        system_prompt = f"""You are the NerdMedica Socratic Auditor, a highly specialized medical tutoring AI. Your purpose is to facilitate clinical reasoning, not just provide answers. 

Follow this structural protocol for EVERY generation:
1. CLINICAL SCENARIO: Create a brief, realistic 3-sentence patient history (Age, Chief Complaint, Vitals) that matches the pathology seen in the provided {modality}.
2. ANATOMICAL INVENTORY: Explicitly focus on {focus}.
3. DIFFERENTIAL REASONING: Mention the primary finding but immediately contrast it with a 'mimic'.
4. SOCRATIC QUESTIONING: Answer the clinician's query Socraticly. Challenge the clinician to justify their diagnosis based on visual evidence.

Tone: Professional, objective, and Socratic. Ensure you provide a complete answer."""

    prompt_content = f"Clinician Question: {message}"
    
    history_text = ""
    if history:
        for msg in history:
            role = "User" if msg["role"] == "user" else "AI"
            content = msg["content"]
            history_text += f"\n{role}: {content}"
            
    if history_text:
        formatted_prompt = f"{system_prompt}\n\nChat History:{history_text}\n\n{prompt_content}"
    else:
        formatted_prompt = f"{system_prompt}\n\n{prompt_content}"
    
    messages = [
        {"role": "user", "content": formatted_prompt}
    ]
    
    print("Generating NerdMedica feedback using MedGemma (Streaming)...")
    try:
        prompt_str = tokenizer.apply_chat_template(
            messages, add_generation_prompt=True, tokenize=False
        )
        inputs = tokenizer(prompt_str, return_tensors="pt")
        inputs = {k: v.to(model.device) for k, v in inputs.items()}
        
        streamer = TextIteratorStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True)
        
        generation_kwargs = dict(
            **inputs,
            streamer=streamer,
            max_new_tokens=512,
            temperature=0.1,
            top_p=0.9,
            repetition_penalty=1.2,
            do_sample=True
        )
        
        thread = Thread(target=model.generate, kwargs=generation_kwargs)
        thread.start()
        
        generated_text = ""
        for new_text in streamer:
            generated_text += new_text
            yield generated_text
        
    except Exception as e:
        print(f"Error during MedGemma generation: {e}")
        yield f"### Generation Error\nSystem Recalibrating: VRAM constraint exceeded or model error encountered. Details: {e}"