src/synthetic_generator.py

import pandas as pd
import numpy as np
import torch
import re
from transformers import AutoTokenizer, AutoModelForCausalLM

def generate_synthetic_case(clinical_query, model_id="meta-llama/Llama-3.2-3B-Instruct", max_tokens=800):
    """Generate a synthetic clinical case with examiner questions and expected answers."""
    print(f"Generating synthetic case for '{clinical_query}' using {model_id}...")
    gen_tokenizer = None
    gen_model = None
    try:
        # Initialize generator model components
        gen_tokenizer = AutoTokenizer.from_pretrained(model_id)
        gen_model = AutoModelForCausalLM.from_pretrained(
            model_id,
            torch_dtype=torch.float16,
            device_map="auto"
        )
        gen_model.eval()
        device = gen_model.device
        if gen_tokenizer.pad_token is None:
            gen_tokenizer.pad_token = gen_tokenizer.eos_token

    except Exception as e:
        print(f"Error initializing generator model {model_id}: {e}")
        return None

    prompt = f"""<s>[INST] You are a board-certified general surgeon simulating a clinical oral board exam.
    Create a synthetic case on the topic: "{clinical_query}".
    Start by describing the initial clinical presentation in 1–2 sentences.
    Then generate a list of 5–8 examiner questions (Q1, Q2...), each paired with the expected examinee answer (A1, A2...). Ensure Q/A pairs are clearly separated.
    Output ONLY the presentation and Q&A pairs in this exact format:
    Clinical Presentation: ...
    
    Q1: ...
    A1: ...
    
    Q2: ...
    A2: ...
    
    (continue until Qn/An)
    Focus on common scenarios and standard knowledge. Avoid overly complex or rare details.
    [/INST]</s>"""

    output_text = None
    try:
        inputs = gen_tokenizer(prompt, return_tensors="pt").to(device)
        input_ids_length = inputs.input_ids.shape[1]

        with torch.no_grad():
            outputs = gen_model.generate(
                inputs.input_ids,
                max_new_tokens=max_tokens,
                do_sample=True, # Sample to get potentially varied outputs
                temperature=0.7,
                top_p=0.9,
                pad_token_id=gen_tokenizer.eos_token_id
            )

        # Decode only the newly generated tokens
        generated_ids = outputs[0][input_ids_length:]
        output_text = gen_tokenizer.decode(generated_ids, skip_special_tokens=True).strip()
        print("Synthetic case generation complete.")

    except Exception as e:
        print(f"Error during synthetic case generation: {e}")
    finally:
        # Clean up model resources
        del gen_model
        del gen_tokenizer
        if torch.cuda.is_available():
             torch.cuda.empty_cache()
        return output_text

def process_synthetic_data(clinical_query, output_text):
    """Process the raw LLM output text into a structured DataFrame for the DummyRetriever."""
    # Extract clinical presentation
    match = re.search(r"Clinical Presentation:(.*?)(?=\n\nQ1:|$)", output_text, re.DOTALL | re.IGNORECASE)
    clinical_presentation_text = match.group(1).strip() if match else "Synthetic Case: " + clinical_query

    # Extract Q&A pairs
    qa_pattern = r"Q(\d+):\s*(.*?)\s*A\1:\s*(.*?)(?=\n*Q\d+:|\Z)"
    qa_matches = re.findall(qa_pattern, output_text, flags=re.DOTALL | re.IGNORECASE)

    qa_list = []
    for match_tuple in qa_matches:
        try:
            q_num = int(match_tuple[0])
            q_text = match_tuple[1].strip()
            a_text = match_tuple[2].strip()
            if q_text and a_text:
                qa_list.append({'turn_id': q_num, 'question': q_text, 'answer': a_text})
        except (IndexError, ValueError) as e:
             print(f"Warning: Skipping malformed Q/A match: {match_tuple} due to {e}")

    if not qa_list:
        print("Warning: No valid Q&A pairs extracted from synthetic text.")
        return pd.DataFrame()

    qa_list.sort(key=lambda item: item['turn_id'])

    rows = []
    for item in qa_list:
        rows.append({
            'case_id': 'SYNTH_01',
            'clinical_presentation': clinical_query, # Use query as presentation title
            'turn_id': item['turn_id'],
            'question': item['question'],
            'answer': item['answer']
        })

    df_synthetic = pd.DataFrame(rows)

    if not df_synthetic.empty and clinical_presentation_text:
         # Find the index of the first turn
         first_turn_index = df_synthetic[df_synthetic['turn_id'] == 1].index
         if not first_turn_index.empty:
              idx = first_turn_index[0]
              df_synthetic.loc[idx, 'question'] = clinical_presentation_text + " " + df_synthetic.loc[idx, 'question']
         else:
              print("Warning: Could not find turn_id 1 to prepend presentation.")

    print(f"Processed synthetic data into DataFrame with {len(df_synthetic)} turns.")
    return df_synthetic