Use this to see *what context a RAG system would surface* for a question, and how that
changes with the retrieval method. The corpus is a fictional B2B SaaS knowledge base that
ships with the server, so no setup or data is required.

Args:
    query: The natural-language question or search string.
    k: How many passages to return (1 to 20).
    method: One of 'bm25', 'tfidf', 'dense', or 'hybrid' (default 'hybrid').

Returns:
    RetrieveResult with fields:
        - query (str), method (str), k (int), count (int)
        - passages: list of {rank, doc_id, chunk_id, score, text}, best first.

Raises:
    A tool error if method='dense' but the optional 'sentence-transformers' extra is not
    installed; the message explains how to enable it.

Runs every labeled question through the chosen retriever and averages four ranking
metrics, so you can put a number on retrieval quality instead of eyeballing it.

Args:
    method: Which retrieval strategy to score (default 'hybrid').
    k: The cutoff for recall@k, precision@k, and nDCG@k (1 to 20, default 3).

Returns:
    MetricsOut with fields: method, k, n_questions, recall_at_k, precision_at_k,
    mrr_at_k, ndcg_at_k. Each metric is in the range 0.0 to 1.0.

Raises:
    A tool error if method='dense' and the optional dense extra is not installed.

The honest, defensible answer to "which retrieval strategy should we use?". Each method
is scored over the full labeled question set and ranked by nDCG@k. Methods whose optional
dependencies are missing (currently only 'dense') are reported under ``skipped`` with a
reason rather than failing the whole call.

Args:
    k: The cutoff applied to recall@k, precision@k, and nDCG@k (1 to 20, default 3).

Returns:
    CompareResult with fields:
        - k (int), n_questions (int)
        - rows: list of {method, recall_at_k, precision_at_k, mrr_at_k, ndcg_at_k}
        - best_method (str): the runnable method with the highest nDCG@k
        - skipped: list of {method, reason} for methods that could not run.