def llm_rerank(
    query,
    candidate_df,
    text_column="child_text",
    top_k=3,
):
    candidates = [
        {
            "candidate_id": i,
            "text": row[text_column],
        }
        for i, (_, row) in enumerate(
            candidate_df.iterrows(),
            start=1,
        )
    ]

    payload = {
        "query": query,
        "instruction": (
            "Rank candidates by how directly "
            "and completely they help answer "
            f"the query. Return the best {top_k} IDs."
        ),
        "candidates": candidates,
    }

    response = client.responses.create(
        model=RERANK_MODEL,
        input=[
            {
                "role": "system",
                "content": (
                    'Return valid JSON only: '
                    '{"ranked_candidate_ids":[1,2,3]}'
                ),
            },
            {
                "role": "user",
                "content": json.dumps(payload),
            },
        ],
    )

    ranked_ids = json.loads(
        response.output_text
    )["ranked_candidate_ids"][:top_k]

    rows = []

    for new_rank, candidate_id in enumerate(
        ranked_ids,
        start=1,
    ):
        if 1 <= candidate_id <= len(candidate_df):
            row = candidate_df.iloc[
                candidate_id - 1
            ].copy()

            row["initial_rank"] = candidate_id
            row["reranked_rank"] = new_rank
            rows.append(row)

    return pd.DataFrame(rows)

        __
        __