mini_ladder.py

from transformers import pipeline, AutoTokenizer

# ------------------------------
# 1) CRITIQUE MODEL & TOKENIZER (for fast mode, this step can be skipped)
# ------------------------------
CRITIQUE_MODEL_NAME = "distilgpt2"
critique_pipeline = pipeline("text-generation", model=CRITIQUE_MODEL_NAME)
critique_tokenizer = AutoTokenizer.from_pretrained(CRITIQUE_MODEL_NAME)

# DistilGPT-2 has a maximum context length of ~1024 tokens.
GPT2_MAX_CONTEXT = 1024

# ------------------------------
# 2) SUB-QUESTION GENERATION
# ------------------------------
def generate_sub_questions(main_query: str):
    """
    Naively generates sub-questions for the given main query.
    """
    return [
        f"1) What are common causes of {main_query}?",
        f"2) Which medications are typically used for {main_query}?",
        f"3) What are non-pharmacological approaches to {main_query}?"
    ]

# ------------------------------
# 3) SELF-CRITIQUE & REFINEMENT
# ------------------------------
def self_critique_and_refine(query: str, initial_answer: str, docs: list, fast_mode: bool = False):
    """
    Uses a smaller model (DistilGPT-2) for self-critique and optionally refines the answer.
    If fast_mode is True, the self-critique step is skipped, and the initial answer is returned.
    """
    if fast_mode:
        # Fast mode: Skip self-critique and refinement
        return initial_answer, "Self-critique skipped for speed."

    # Construct the critique prompt.
    critique_prompt = (
        f"The following is an answer to the question '{query}'. "
        "Evaluate its correctness, clarity, and completeness. "
        "List any missing details or inaccuracies.\n\n"
        f"ANSWER:\n{initial_answer}\n\n"
        "CRITIQUE:"
    )
    
    # Truncate the prompt to ensure there's room for 20 new tokens.
    truncated_prompt = _truncate_prompt_for_gpt2(critique_prompt, buffer=20)
    
    # Generate the critique using DistilGPT-2.
    critique_gen = critique_pipeline(
        truncated_prompt,
        max_new_tokens=20,
        truncation=True
    )
    if critique_gen and isinstance(critique_gen, list):
        critique_text = critique_gen[0]["generated_text"]
    else:
        critique_text = "No critique generated."
    
    # If critique flags issues, refine the answer using BioGPT.
    if any(word in critique_text.lower() for word in ["missing", "incomplete", "incorrect", "lacks"]):
        refine_prompt = (
            f"Question: {query}\n"
            f"Current Answer: {initial_answer}\n"
            f"Critique: {critique_text}\n"
            "Refine the answer by adding missing or corrected information. "
            "Use the context below if needed:\n\n"
            + "\n\n".join(docs)
            + "\nREFINED ANSWER:"
        )
        from backend import qa_pipeline  # Import here to avoid circular imports.
        refined_gen = qa_pipeline(refine_prompt, max_new_tokens=120, truncation=True)
        if refined_gen and isinstance(refined_gen, list):
            refined_answer = refined_gen[0]["generated_text"]
        else:
            refined_answer = initial_answer
    else:
        refined_answer = initial_answer

    return refined_answer, critique_text

# ------------------------------
# 4) HELPER: GPT-2 PROMPT TRUNCATION
# ------------------------------
def _truncate_prompt_for_gpt2(prompt_text: str, buffer: int = 20) -> str:
    """
    Truncates the input prompt so that its token count plus a reserved buffer does not exceed GPT-2's maximum context.
    """
    tokens = critique_tokenizer.encode(prompt_text, add_special_tokens=False)
    max_allowed = GPT2_MAX_CONTEXT - buffer
    if len(tokens) > max_allowed:
        tokens = tokens[:max_allowed]
    truncated_text = critique_tokenizer.decode(tokens, skip_special_tokens=True)
    return truncated_text