#     similarity_matrix = spec2vec_similarity.matrix(
    #         uniq_documents_processed, uniq_documents_processed,
    #         is_symmetric=True)
    #     np.save(sims_out, similarity_matrix)
    # else:
    #     similarity_matrix = np.load(sims_out)

    # classical mod cosine similarities
    mod_cos_sims_out = os.path.join(
        cmd.output_dir, "similarities_unique_inchikey_mod_cosine.npy")
    if not os.path.exists(mod_cos_sims_out):
        similarity_measure = ModifiedCosine(tolerance=0.005,
                                            mz_power=0,
                                            intensity_power=1.0)
        mod_cos_similarity = similarity_measure.matrix(
            uniq_spectrums_classical,
            uniq_spectrums_classical,
            is_symmetric=True)
        np.save(mod_cos_sims_out, mod_cos_similarity)
    else:
        mod_cos_similarity = np.load(mod_cos_sims_out)

    # # md s2v similarities
    # md_sims_out = os.path.join(
    #     cmd.output_dir,
    #     'similarities_unique_inchikey_mds_spec2vec_librarymodel.npy')
    # if not os.path.exists(md_sims_out):
    #     md_spec2vec_similarity = Spec2Vec(model_mds,
    #                                       intensity_weighting_power=0.5)
    #     md_similarity_matrix = md_spec2vec_similarity.matrix(
    #         uniq_documents_mds, uniq_documents_mds, is_symmetric=True)
    #     np.save(md_sims_out, md_similarity_matrix)
def library_matching(documents_query: List[SpectrumDocument],
                     documents_library: List[SpectrumDocument],
                     model,
                     presearch_based_on=["parentmass", "spec2vec-top10"],
                     ignore_non_annotated: bool = True,
                     include_scores=["spec2vec", "cosine", "modcosine"],
                     intensity_weighting_power: float = 0.5,
                     allowed_missing_percentage: float = 0,
                     cosine_tol: float = 0.005,
                     mass_tolerance: float = 1.0):
    """Selecting potential spectra matches with spectra library.

    Suitable candidates will be selected by 1) top_n Spec2Vec similarity, and 2)
    same precursor mass (within given mz_ppm tolerance(s)).
    For later matching routines, additional scores (cosine, modified cosine)
    are added as well.

    Args:
    --------
    documents_query:
        List containing all spectrum documents that should be queried against the library.
    documents_library:
        List containing all library spectrum documents.
    model:
        Pretrained word2Vec model.
    top_n: int, optional
        Number of entries witht the top_n highest Spec2Vec scores to keep as
        found matches. Default = 10.
    ignore_non_annotated: bool, optional
        If True, only annotated spectra will be considered for matching.
        Default = True.
    cosine_tol: float, optional
        Set tolerance for the cosine and modified cosine score. Default = 0.005
    mass_tolerance
        Specify tolerance for a parentmass match.
    """

    # Initializations
    found_matches = []
    m_mass_matches = None
    m_spec2vec_similarities = None

    def get_metadata(documents):
        metadata = []
        for doc in documents:
            metadata.append(doc._obj.get("smiles"))
        return metadata

    library_spectra_metadata = get_metadata(documents_library)
    if ignore_non_annotated:
        # Get array of all ids for spectra with smiles
        library_ids = np.asarray(
            [i for i, x in enumerate(library_spectra_metadata) if x])
    else:
        library_ids = np.arange(len(documents_library))

    msg = "Presearch must be done either by 'parentmass' and/or 'spec2vec-topX'"
    assert "parentmass" in presearch_based_on or np.any(
        ["spec2vec" in x for x in presearch_based_on]), msg

    # 1. Search for top-n Spec2Vec matches ------------------------------------
    if np.any(["spec2vec" in x for x in presearch_based_on]):
        top_n = int([
            x.split("top")[1] for x in presearch_based_on if "spec2vec" in x
        ][0])
        print("Pre-selection includes spec2vec top {}.".format(top_n))
        spec2vec = Spec2Vec(
            model=model,
            intensity_weighting_power=intensity_weighting_power,
            allowed_missing_percentage=allowed_missing_percentage)
        m_spec2vec_similarities = spec2vec.matrix(
            [documents_library[i] for i in library_ids], documents_query)

        # Select top_n similarity values:
        selection_spec2vec = np.argpartition(m_spec2vec_similarities,
                                             -top_n,
                                             axis=0)[-top_n:, :]
    else:
        selection_spec2vec = np.empty((0, len(documents_query)), dtype="int")

    # 2. Search for parent mass based matches ---------------------------------
    if "parentmass" in presearch_based_on:
        mass_matching = ParentmassMatch(mass_tolerance)
        m_mass_matches = mass_matching.matrix(
            [documents_library[i]._obj for i in library_ids],
            [x._obj for x in documents_query])
        selection_massmatch = []
        for i in range(len(documents_query)):
            selection_massmatch.append(np.where(m_mass_matches[:, i] == 1)[0])
    else:
        selection_massmatch = np.empty((len(documents_query), 0), dtype="int")

    # 3. Combine found matches ------------------------------------------------
    for i in range(len(documents_query)):
        s2v_top_ids = selection_spec2vec[:, i]
        mass_match_ids = selection_massmatch[i]

        all_match_ids = np.unique(np.concatenate(
            (s2v_top_ids, mass_match_ids)))

        if len(all_match_ids) > 0:
            if "modcosine" in include_scores:
                # Get cosine score for found matches
                cosine_similarity = CosineGreedy(tolerance=cosine_tol)
                cosine_scores = []
                for match_id in library_ids[all_match_ids]:
                    cosine_scores.append(
                        cosine_similarity.matrix(
                            documents_library[match_id]._obj,
                            documents_query[i]._obj))
            else:
                cosine_scores = len(all_match_ids) * ["not calculated"]

            if "cosine" in include_scores:
                # Get modified cosine score for found matches
                mod_cosine_similarity = ModifiedCosine(tolerance=cosine_tol)
                mod_cosine_scores = []
                for match_id in library_ids[all_match_ids]:
                    mod_cosine_scores.append(
                        mod_cosine_similarity.matrix(
                            documents_library[match_id]._obj,
                            documents_query[i]._obj))
            else:
                mod_cosine_scores = len(all_match_ids) * ["not calculated"]

            matches_df = pd.DataFrame(
                {
                    "cosine_score": [x[0] for x in cosine_scores],
                    "cosine_matches": [x[1] for x in cosine_scores],
                    "mod_cosine_score": [x[0] for x in mod_cosine_scores],
                    "mod_cosine_matches": [x[1] for x in mod_cosine_scores]
                },
                index=library_ids[all_match_ids])

            if m_mass_matches is not None:
                matches_df["mass_match"] = m_mass_matches[all_match_ids, i]

            if m_spec2vec_similarities is not None:
                matches_df["s2v_score"] = m_spec2vec_similarities[
                    all_match_ids, i]
            elif "spec2vec" in include_scores:
                spec2vec_similarity = Spec2Vec(
                    model=model,
                    intensity_weighting_power=intensity_weighting_power,
                    allowed_missing_percentage=allowed_missing_percentage)
                spec2vec_scores = []
                for match_id in library_ids[all_match_ids]:
                    spec2vec_scores.append(
                        spec2vec_similarity.pair(documents_library[match_id],
                                                 documents_query[i]))
                matches_df["s2v_score"] = spec2vec_scores
            found_matches.append(matches_df.fillna(0))
        else:
            found_matches.append([])

    return found_matches