models/distilbert_model.py

"""
distilbert_model.py — HuggingFace DistilBERT Question Answering Model.

Model: distilbert-base-cased-distilled-squad
40% smaller and 60% faster than BERT with ~97% of its performance.
Uses direct PyTorch inference (compatible with transformers 5.x).
"""

import logging

logger = logging.getLogger(__name__)

_tokenizer = None
_model = None
MODEL_NAME = "distilbert-base-cased-distilled-squad"


def init_distilbert_model():
    """Load the DistilBERT QA model. Called once at app startup."""
    global _tokenizer, _model
    try:
        from transformers import AutoTokenizer, AutoModelForQuestionAnswering
        logger.info(f"[DistilBERT] Loading model '{MODEL_NAME}' ...")
        _tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
        _model = AutoModelForQuestionAnswering.from_pretrained(MODEL_NAME)
        _model.eval()
        logger.info("[DistilBERT] Model loaded and ready.")
    except Exception as exc:
        logger.error(f"[DistilBERT] Failed to load model: {exc}")
        _tokenizer = None
        _model = None


def _run_qa_inference(context: str, question: str) -> dict:
    """Direct PyTorch inference — works with any transformers version."""
    import torch
    import torch.nn.functional as F

    inputs = _tokenizer(
        question, context,
        return_tensors="pt",
        truncation=True,
        max_length=512,
    )

    with torch.no_grad():
        outputs = _model(**inputs)

    start_logits = outputs.start_logits[0]
    end_logits   = outputs.end_logits[0]

    start_idx = int(torch.argmax(start_logits))
    end_idx   = int(torch.argmax(end_logits)) + 1

    if end_idx <= start_idx:
        end_idx = start_idx + 1

    input_ids = inputs["input_ids"][0]
    answer_tokens = input_ids[start_idx:end_idx]
    answer = _tokenizer.decode(answer_tokens, skip_special_tokens=True).strip()

    # DistilBERT Normalized Calibration: Softer scaling to respect BERT-Large's superiority.
    import math
    start_prob = float(F.softmax(start_logits, dim=0)[start_idx])
    end_prob   = float(F.softmax(end_logits,   dim=0)[end_idx - 1])
    avg_prob = (start_prob + end_prob) / 2
    calibrated_score = avg_prob ** 0.8  # Softer scaling
    score = round(min(max(calibrated_score, 0.0), 1.0), 4)

    # --- Strong Collision Filter: Block question repetition using word overlap ---
    q_words = set(question.lower().replace('?', '').split())
    a_words = set(answer.lower().replace('?', '').split())
    
    if q_words and a_words:
        common = q_words.intersection(a_words)
        overlap_ratio = len(common) / len(q_words)
    else:
        overlap_ratio = 0

    if overlap_ratio > 0.7 or len(answer.strip()) < 1 or "[cls]" in answer.lower():
        return {
            "answer": "I'm sorry, I couldn't find a specific answer to that in the provided document.",
            "score": 0.0,
            "error": False,
            "not_found": True
        }

    return {"answer": answer, "score": score}


def predict(context: str, question: str) -> dict:
    """
    Run QA inference.

    Returns:
        {
            "answer": str,
            "score": float (0.0–1.0),
            "model": "DistilBERT",
            "model_id": "distilbert"
        }
    """
    if _model is None or _tokenizer is None:
        return {
            "answer": "DistilBERT model is not loaded. Please check server logs.",
            "score": 0.0,
            "model": "DistilBERT",
            "model_id": "distilbert",
            "error": True,
        }

    if not context or not question:
        return {
            "answer": "Context and question must not be empty.",
            "score": 0.0,
            "model": "DistilBERT",
            "model_id": "distilbert",
            "error": True,
        }

    try:
        result = _run_qa_inference(context=context, question=question)
        score = result["score"]
        answer = result["answer"]

        if score < 0.05 or "[CLS]" in answer or not answer:
            answer = "Answer not found with sufficient confidence. Try rephrasing your question or providing more context."
            score = 0.0

        return {
            "answer": answer,
            "score": score,
            "model": "DistilBERT",
            "model_id": "distilbert",
            "error": False,
        }
    except Exception as exc:
        logger.error(f"[DistilBERT] Inference error: {exc}")
        return {
            "answer": f"Inference error: {exc}",
            "score": 0.0,
            "model": "DistilBERT",
            "model_id": "distilbert",
            "error": True,
        }