mteb 2.5.2__py3-none-any.whl → 2.5.4__py3-none-any.whl

mteb/models/model_implementations/kennethenevoldsen_models.py

@@ -4,7 +4,7 @@ from mteb.models.sentence_transformer_wrapper import (
 )
 
 dfm_enc_large = ModelMeta(
-    loader=sentence_transformers_loader,  # type: ignore
+    loader=sentence_transformers_loader,
     name="KennethEnevoldsen/dfm-sentence-encoder-large",
     model_type=["dense"],
     languages=["dan-Latn"],
@@ -39,7 +39,7 @@ dfm_enc_large = ModelMeta(
 )
 
 dfm_enc_med = ModelMeta(
-    loader=sentence_transformers_loader,  # type: ignore
+    loader=sentence_transformers_loader,
     name="KennethEnevoldsen/dfm-sentence-encoder-medium",
     model_type=["dense"],
     languages=["dan-Latn"],

mteb/models/model_implementations/llm2clip_models.py

@@ -181,7 +181,7 @@ llm2clip_training_sets = set(
 )
 
 llm2clip_openai_l_14_336 = ModelMeta(
-    loader=llm2clip_loader,  # type: ignore
+    loader=llm2clip_loader,
     name="microsoft/LLM2CLIP-Openai-L-14-336",
     model_type=["dense"],
     languages=["eng-Latn"],
@@ -206,7 +206,7 @@ llm2clip_openai_l_14_336 = ModelMeta(
 
 # NOTE: https://huggingface.co/microsoft/LLM2CLIP-Openai-L-14-224/discussions/1
 llm2clip_openai_l_14_224 = ModelMeta(
-    loader=llm2clip_loader,  # type: ignore
+    loader=llm2clip_loader,
     name="microsoft/LLM2CLIP-Openai-L-14-224",
     model_type=["dense"],
     languages=["eng-Latn"],
@@ -230,7 +230,7 @@ llm2clip_openai_l_14_224 = ModelMeta(
 )
 
 llm2clip_openai_b_16 = ModelMeta(
-    loader=llm2clip_loader,  # type: ignore
+    loader=llm2clip_loader,
     name="microsoft/LLM2CLIP-Openai-B-16",
     model_type=["dense"],
     languages=["eng-Latn"],

mteb/models/model_implementations/mcinext_models.py

@@ -1,6 +1,7 @@
 import logging
 import os
 import time
+import warnings
 from typing import Any
 
 import numpy as np
@@ -246,7 +247,9 @@ class HakimModelWrapper(AbsEncoder):
         task_prompt, task_id = DATASET_TASKS.get(task_name, (None, None))
 
         if not task_prompt:
-            logger.warning(f"Unknown dataset: {task_name}, no preprocessing applied.")
+            msg = f"Unknown dataset: {task_name}, no preprocessing applied."
+            logger.warning(msg)
+            warnings.warn(msg)
             return sample
 
         task_prompt = f"مسئله : {task_prompt}"

mteb/models/model_implementations/moco_models.py

@@ -117,7 +117,7 @@ mocov3_training_datasets = set(
 )
 
 mocov3_vit_base = ModelMeta(
-    loader=mocov3_loader,  # type: ignore
+    loader=mocov3_loader,
     name="nyu-visionx/moco-v3-vit-b",
     model_type=["dense"],
     languages=["eng-Latn"],
@@ -141,7 +141,7 @@ mocov3_vit_base = ModelMeta(
 )
 
 mocov3_vit_large = ModelMeta(
-    loader=mocov3_loader,  # type: ignore
+    loader=mocov3_loader,
     name="nyu-visionx/moco-v3-vit-l",
     model_type=["dense"],
     languages=["eng-Latn"],

mteb/models/model_implementations/model2vec_models.py

@@ -139,7 +139,7 @@ class Model2VecModel(AbsEncoder):
             **kwargs: Additional arguments to pass to the wrapper.
         """
         requires_package(self, "model2vec", model_name, "pip install 'mteb[model2vec]'")
-        from model2vec import StaticModel  # type: ignore
+        from model2vec import StaticModel
 
         self.model_name = model_name
         self.model = StaticModel.from_pretrained(self.model_name)

mteb/models/model_implementations/nomic_models.py

@@ -193,7 +193,7 @@ NOMIC_CITATION = """
 """
 
 nomic_embed_v1_5 = ModelMeta(
-    loader=NomicWrapper,  # type: ignore
+    loader=NomicWrapper,
     loader_kwargs=dict(
         trust_remote_code=True,
         model_prompts=model_prompts,
@@ -222,7 +222,7 @@ nomic_embed_v1_5 = ModelMeta(
 )
 
 nomic_embed_v1 = ModelMeta(
-    loader=NomicWrapper,  # type: ignore
+    loader=NomicWrapper,
     loader_kwargs=dict(
         trust_remote_code=True,
         model_prompts=model_prompts,
@@ -251,7 +251,7 @@ nomic_embed_v1 = ModelMeta(
 )
 
 nomic_embed_v1_ablated = ModelMeta(
-    loader=NomicWrapper,  # type: ignore
+    loader=NomicWrapper,
     loader_kwargs=dict(
         trust_remote_code=True,
         model_prompts=model_prompts,
@@ -279,7 +279,7 @@ nomic_embed_v1_ablated = ModelMeta(
 )
 
 nomic_embed_v1_unsupervised = ModelMeta(
-    loader=NomicWrapper,  # type: ignore
+    loader=NomicWrapper,
     loader_kwargs=dict(
         trust_remote_code=True,
         model_prompts=model_prompts,
@@ -334,7 +334,7 @@ nomic_modern_bert_embed = ModelMeta(
     training_datasets=nomic_training_data,
     public_training_data=None,
     citation="""@misc{nussbaum2024nomic,
-      title={Nomic Embed: Training a Reproducible Long Context Text Embedder}, 
+      title={Nomic Embed: Training a Reproducible Long Context Text Embedder},
       author={Zach Nussbaum and John X. Morris and Brandon Duderstadt and Andriy Mulyar},
       year={2024},
       eprint={2402.01613},
@@ -446,7 +446,7 @@ m_languages = [
 ]
 
 nomic_embed_text_v2_moe = ModelMeta(
-    loader=NomicWrapper,  # type: ignore
+    loader=NomicWrapper,
     loader_kwargs=dict(
         trust_remote_code=True,
         model_prompts=model_prompts,
@@ -472,12 +472,12 @@ nomic_embed_text_v2_moe = ModelMeta(
     training_datasets=None,  # did not look into this further
     superseded_by=None,
     citation="""@misc{nussbaum2025trainingsparsemixtureexperts,
-      title={Training Sparse Mixture Of Experts Text Embedding Models}, 
+      title={Training Sparse Mixture Of Experts Text Embedding Models},
       author={Zach Nussbaum and Brandon Duderstadt},
       year={2025},
       eprint={2502.07972},
       archivePrefix={arXiv},
       primaryClass={cs.CL},
-      url={https://arxiv.org/abs/2502.07972}, 
+      url={https://arxiv.org/abs/2502.07972},
 }""",
 )

mteb/models/model_implementations/openclip_models.py

@@ -120,7 +120,7 @@ def openclip_loader(model_name, **kwargs):
 
 
 CLIP_ViT_L_14_DataComp_XL_s13B_b90K = ModelMeta(
-    loader=openclip_loader,  # type: ignore
+    loader=openclip_loader,
     name="laion/CLIP-ViT-L-14-DataComp.XL-s13B-b90K",
     model_type=["dense"],
     languages=["eng-Latn"],
@@ -146,7 +146,7 @@ CLIP_ViT_L_14_DataComp_XL_s13B_b90K = ModelMeta(
 )
 
 CLIP_ViT_B_32_DataComp_XL_s13B_b90K = ModelMeta(
-    loader=openclip_loader,  # type: ignore
+    loader=openclip_loader,
     name="laion/CLIP-ViT-B-32-DataComp.XL-s13B-b90K",
     model_type=["dense"],
     languages=["eng-Latn"],
@@ -172,7 +172,7 @@ CLIP_ViT_B_32_DataComp_XL_s13B_b90K = ModelMeta(
 )
 
 CLIP_ViT_B_16_DataComp_XL_s13B_b90K = ModelMeta(
-    loader=openclip_loader,  # type: ignore
+    loader=openclip_loader,
     name="laion/CLIP-ViT-B-16-DataComp.XL-s13B-b90K",
     model_type=["dense"],
     languages=["eng-Latn"],
@@ -198,7 +198,7 @@ CLIP_ViT_B_16_DataComp_XL_s13B_b90K = ModelMeta(
 )
 
 CLIP_ViT_bigG_14_laion2B_39B_b160k = ModelMeta(
-    loader=openclip_loader,  # type: ignore
+    loader=openclip_loader,
     name="laion/CLIP-ViT-bigG-14-laion2B-39B-b160k",
     model_type=["dense"],
     languages=["eng-Latn"],
@@ -224,7 +224,7 @@ CLIP_ViT_bigG_14_laion2B_39B_b160k = ModelMeta(
 )
 
 CLIP_ViT_g_14_laion2B_s34B_b88K = ModelMeta(
-    loader=openclip_loader,  # type: ignore
+    loader=openclip_loader,
     name="laion/CLIP-ViT-g-14-laion2B-s34B-b88K",
     model_type=["dense"],
     languages=["eng-Latn"],
@@ -250,7 +250,7 @@ CLIP_ViT_g_14_laion2B_s34B_b88K = ModelMeta(
 )
 
 CLIP_ViT_H_14_laion2B_s32B_b79K = ModelMeta(
-    loader=openclip_loader,  # type: ignore
+    loader=openclip_loader,
     name="laion/CLIP-ViT-H-14-laion2B-s32B-b79K",
     model_type=["dense"],
     languages=["eng-Latn"],
@@ -276,7 +276,7 @@ CLIP_ViT_H_14_laion2B_s32B_b79K = ModelMeta(
 )
 
 CLIP_ViT_L_14_laion2B_s32B_b82K = ModelMeta(
-    loader=openclip_loader,  # type: ignore
+    loader=openclip_loader,
     name="laion/CLIP-ViT-L-14-laion2B-s32B-b82K",
     model_type=["dense"],
     languages=["eng-Latn"],

mteb/models/model_implementations/random_baseline.py

@@ -68,7 +68,7 @@ _common_mock_metadata = dict(
     license="mit",
     max_tokens=np.inf,
     reference=None,
-    similarity_fn_name="cosine",  # type: ignore
+    similarity_fn_name="cosine",
     framework=[],
     use_instructions=False,
     public_training_code=None,  # No training code, as this is a random baseline
@@ -187,7 +187,7 @@ class RandomEncoderBaseline:
 
 
 random_encoder_baseline = ModelMeta(
-    loader=RandomEncoderBaseline,  # type: ignore
+    loader=RandomEncoderBaseline,
     name="baseline/random-encoder-baseline",
     model_type=["dense"],
     modalities=["text", "image"],
@@ -232,7 +232,7 @@ class RandomCrossEncoderBaseline:
 
 
 random_cross_encoder_baseline = ModelMeta(
-    loader=RandomCrossEncoderBaseline,  # type: ignore
+    loader=RandomCrossEncoderBaseline,
     name="baseline/random-cross-encoder-baseline",
     model_type=["cross-encoder"],
     modalities=["text", "image"],

mteb/models/model_implementations/rasgaard_models.py

@@ -4,7 +4,7 @@ from mteb.models.model_implementations.model2vec_models import Model2VecModel
 from mteb.models.model_meta import ModelMeta, ScoringFunction
 
 potion_base_8m = ModelMeta(
-    loader=Model2VecModel,  # type: ignore
+    loader=Model2VecModel,
     name="rasgaard/m2v-dfm-large",
     model_type=["dense"],
     languages=["dan-Latn"],

mteb/models/model_implementations/repllama_models.py

@@ -154,7 +154,7 @@ REPLLAMA_CITATION = """
 """
 
 repllama_llama2_original = ModelMeta(
-    loader=RepLLaMAModel,  # type: ignore
+    loader=RepLLaMAModel,
     loader_kwargs=dict(
         base_model_name_or_path="meta-llama/Llama-2-7b-hf",
         device_map="auto",
@@ -187,7 +187,7 @@ repllama_llama2_original = ModelMeta(
 
 
 repllama_llama2_reproduced = ModelMeta(
-    loader=RepLLaMAModel,  # type: ignore
+    loader=RepLLaMAModel,
     loader_kwargs=dict(
         base_model_name_or_path="meta-llama/Llama-2-7b-hf",
         device_map="auto",

mteb/models/model_implementations/rerankers_custom.py

@@ -214,7 +214,7 @@ class JinaReranker(RerankerWrapper):
 
 
 monobert_large = ModelMeta(
-    loader=MonoBERTReranker,  # type: ignore
+    loader=MonoBERTReranker,
     loader_kwargs=dict(
         fp_options="float16",
     ),
@@ -239,7 +239,7 @@ monobert_large = ModelMeta(
 
 # languages unclear: https://huggingface.co/jinaai/jina-reranker-v2-base-multilingual/discussions/28
 jina_reranker_multilingual = ModelMeta(
-    loader=JinaReranker,  # type: ignore
+    loader=JinaReranker,
     loader_kwargs=dict(
         fp_options="float16",
     ),
@@ -263,7 +263,7 @@ jina_reranker_multilingual = ModelMeta(
 )
 
 bge_reranker_v2_m3 = ModelMeta(
-    loader=BGEReranker,  # type: ignore
+    loader=BGEReranker,
     loader_kwargs=dict(
         fp_options="float16",
     ),

mteb/models/model_implementations/rerankers_monot5_based.py

@@ -343,7 +343,7 @@ monot5_small = ModelMeta(
 )
 
 monot5_base = ModelMeta(
-    loader=MonoT5Reranker,  # type: ignore
+    loader=MonoT5Reranker,
     loader_kwargs=dict(
         fp_options="float16",
     ),
@@ -442,7 +442,7 @@ monot5_3b = ModelMeta(
 )
 
 flant5_base = ModelMeta(
-    loader=FLANT5Reranker,  # type: ignore
+    loader=FLANT5Reranker,
     loader_kwargs=dict(
         fp_options="float16",
     ),
@@ -902,7 +902,7 @@ mt5_base_mmarco_v2 = ModelMeta(
 )
 
 mt5_13b_mmarco_100k = ModelMeta(
-    loader=MonoT5Reranker,  # type: ignore
+    loader=MonoT5Reranker,
     loader_kwargs=dict(
         fp_options="float16",
     ),

mteb/models/model_implementations/siglip_models.py

@@ -123,7 +123,7 @@ siglip_training_datasets = set(
 )
 
 siglip_so400m_patch14_224 = ModelMeta(
-    loader=SiglipModelWrapper,  # type: ignore
+    loader=SiglipModelWrapper,
     name="google/siglip-so400m-patch14-224",
     model_type=["dense"],
     languages=["eng-Latn"],
@@ -147,7 +147,7 @@ siglip_so400m_patch14_224 = ModelMeta(
 )
 
 siglip_so400m_patch14_384 = ModelMeta(
-    loader=SiglipModelWrapper,  # type: ignore
+    loader=SiglipModelWrapper,
     name="google/siglip-so400m-patch14-384",
     model_type=["dense"],
     languages=["eng-Latn"],
@@ -171,7 +171,7 @@ siglip_so400m_patch14_384 = ModelMeta(
 )
 
 siglip_so400m_patch16_256_i18n = ModelMeta(
-    loader=SiglipModelWrapper,  # type: ignore
+    loader=SiglipModelWrapper,
     name="google/siglip-so400m-patch16-256-i18n",
     model_type=["dense"],
     languages=["eng-Latn"],
@@ -195,7 +195,7 @@ siglip_so400m_patch16_256_i18n = ModelMeta(
 )
 
 siglip_base_patch16_256_multilingual = ModelMeta(
-    loader=SiglipModelWrapper,  # type: ignore
+    loader=SiglipModelWrapper,
     name="google/siglip-base-patch16-256-multilingual",
     model_type=["dense"],
     languages=["eng-Latn"],
@@ -219,7 +219,7 @@ siglip_base_patch16_256_multilingual = ModelMeta(
 )
 
 siglip_base_patch16_256 = ModelMeta(
-    loader=SiglipModelWrapper,  # type: ignore
+    loader=SiglipModelWrapper,
     name="google/siglip-base-patch16-256",
     model_type=["dense"],
     languages=["eng-Latn"],
@@ -243,7 +243,7 @@ siglip_base_patch16_256 = ModelMeta(
 )
 
 siglip_base_patch16_512 = ModelMeta(
-    loader=SiglipModelWrapper,  # type: ignore
+    loader=SiglipModelWrapper,
     name="google/siglip-base-patch16-512",
     model_type=["dense"],
     languages=["eng-Latn"],
@@ -267,7 +267,7 @@ siglip_base_patch16_512 = ModelMeta(
 )
 
 siglip_base_patch16_384 = ModelMeta(
-    loader=SiglipModelWrapper,  # type: ignore
+    loader=SiglipModelWrapper,
     name="google/siglip-base-patch16-384",
     model_type=["dense"],
     languages=["eng-Latn"],
@@ -291,7 +291,7 @@ siglip_base_patch16_384 = ModelMeta(
 )
 
 siglip_base_patch16_224 = ModelMeta(
-    loader=SiglipModelWrapper,  # type: ignore
+    loader=SiglipModelWrapper,
     name="google/siglip-base-patch16-224",
     model_type=["dense"],
     languages=["eng-Latn"],
@@ -315,7 +315,7 @@ siglip_base_patch16_224 = ModelMeta(
 )
 
 siglip_large_patch16_256 = ModelMeta(
-    loader=SiglipModelWrapper,  # type: ignore
+    loader=SiglipModelWrapper,
     name="google/siglip-large-patch16-256",
     model_type=["dense"],
     languages=["eng-Latn"],
@@ -339,7 +339,7 @@ siglip_large_patch16_256 = ModelMeta(
 )
 
 siglip_large_patch16_384 = ModelMeta(
-    loader=SiglipModelWrapper,  # type: ignore
+    loader=SiglipModelWrapper,
     name="google/siglip-large-patch16-384",
     model_type=["dense"],
     languages=["eng-Latn"],

mteb/models/model_implementations/vlm2vec_models.py

@@ -41,7 +41,7 @@ class VLM2VecWrapper(AbsEncoder):
             model_name,
             "pip install flash-attn --no-build-isolation",
         ):
-            import flash_attn  # noqa
+            pass
 
         requires_package(self, "peft", model_name, "pip install 'mteb[peft]'")
         from peft import LoraConfig, PeftModel

mteb/models/model_implementations/voyage_v.py

@@ -40,15 +40,15 @@ def _downsample_image(
         logging.info(
             f"Downsampling image from {width}x{height} to {new_width}x{new_height}"
         )
-        return image.resize(new_size, Image.LANCZOS)  # type: ignore
+        return image.resize(new_size, Image.LANCZOS)
     if width > height:
         if width > 10000:
             logging.error("Processing extremely wide images.")
-            return image.resize((10000, height), Image.LANCZOS)  # type: ignore
+            return image.resize((10000, height), Image.LANCZOS)
     else:
         if height > 10000:
             logging.error("Processing extremely high images.")
-            return image.resize((width, 10000), Image.LANCZOS)  # type: ignore
+            return image.resize((width, 10000), Image.LANCZOS)
     return image
 
 
@@ -202,7 +202,7 @@ def voyage_v_loader(model_name, **kwargs):
 
 
 voyage_v = ModelMeta(
-    loader=voyage_v_loader,  # type: ignore
+    loader=voyage_v_loader,
     name="voyageai/voyage-multimodal-3",
     model_type=["dense"],
     languages=[],  # Unknown

mteb/models/model_meta.py

@@ -81,7 +81,7 @@ def _get_loader_name(
     return loader.__name__
 
 
-_SENTENCE_TRANSFORMER_LIB_NAME = "Sentence Transformers"
+_SENTENCE_TRANSFORMER_LIB_NAME: FRAMEWORKS = "Sentence Transformers"
 
 
 class ModelMeta(BaseModel):
@@ -263,10 +263,8 @@ class ModelMeta(BaseModel):
         _kwargs = self.loader_kwargs.copy()
         _kwargs.update(kwargs)
 
-        model: EncoderProtocol = self.loader(
-            self.name, revision=self.revision, **_kwargs
-        )
-        model.mteb_model_meta = self  # type: ignore
+        model: MTEBModels = self.loader(self.name, revision=self.revision, **_kwargs)
+        model.mteb_model_meta = self  # type: ignore[misc]
         return model
 
     def model_name_as_path(self) -> str:
@@ -318,9 +316,8 @@ class ModelMeta(BaseModel):
                 model_config = None
                 logger.warning(f"Can't get configuration for {model_name}. Error: {e}")
 
-            if (
-                card_data.library_name == _SENTENCE_TRANSFORMER_LIB_NAME
-                or _SENTENCE_TRANSFORMER_LIB_NAME in card_data.tags
+            if card_data.library_name == _SENTENCE_TRANSFORMER_LIB_NAME or (
+                card_data.tags and _SENTENCE_TRANSFORMER_LIB_NAME in card_data.tags
             ):
                 frameworks.append(_SENTENCE_TRANSFORMER_LIB_NAME)
             else:
@@ -435,7 +432,7 @@ class ModelMeta(BaseModel):
                 and config_sbert.get("similarity_fn_name") is not None
             ):
                 meta.similarity_fn_name = ScoringFunction.from_str(
-                    config_sbert.get("similarity_fn_name")
+                    config_sbert["similarity_fn_name"]
                 )
             else:
                 meta.similarity_fn_name = ScoringFunction.COSINE
@@ -511,10 +508,12 @@ class ModelMeta(BaseModel):
                 if adapted_training_datasets is not None:
                     training_datasets |= adapted_training_datasets
             except (ValueError, KeyError) as e:
-                logger.warning(f"Could not get source model: {e} in MTEB")
+                msg = f"Could not get source model: {e} in MTEB"
+                logger.warning(msg)
+                warnings.warn(msg)
 
         return_dataset = training_datasets.copy()
-        visited = set()
+        visited: set[str] = set()
 
         for dataset in training_datasets:
             similar_tasks = _collect_similar_tasks(dataset, visited)
@@ -548,6 +547,8 @@ class ModelMeta(BaseModel):
 
     @staticmethod
     def _calculate_num_parameters_from_hub(model_name: str | None = None) -> int | None:
+        if not model_name:
+            return None
         try:
             safetensors_metadata = get_safetensors_metadata(model_name)
             if len(safetensors_metadata.parameter_count) >= 0:
@@ -561,7 +562,7 @@ class ModelMeta(BaseModel):
             logger.warning(
                 f"Can't calculate number of parameters for {model_name}. Got error {e}"
             )
-            return None
+        return None
 
     def calculate_num_parameters_from_hub(self) -> int | None:
         """Calculates the number of parameters in the model.
@@ -624,7 +625,7 @@ class ModelMeta(BaseModel):
         if "API" in self.framework or self.name is None:
             return None
 
-        return self._calculate_memory_usage_mb(self.model_name, self.n_parameters)
+        return self._calculate_memory_usage_mb(self.name, self.n_parameters)
 
     @staticmethod
     def fetch_release_date(model_name: str) -> StrDate | None:

mteb/models/search_encoder_index/search_indexes/faiss_search_index.py

@@ -1,4 +1,5 @@
 import logging
+import warnings
 from collections.abc import Callable
 
 import numpy as np
@@ -108,7 +109,7 @@ class FaissSearchIndex:
             ids = ids.tolist()
 
         if issubclass(self.index_type, faiss.IndexFlatL2):
-            similarities = -np.sqrt(np.maximum(similarities, 0))
+            similarities = (-np.sqrt(np.maximum(similarities, 0))).tolist()
 
         return similarities, ids
 
@@ -116,8 +117,8 @@ class FaissSearchIndex:
         self,
         embeddings: Array,
         top_k: int,
-        top_ranked: TopRankedDocumentsType | None = None,
-        query_idx_to_id: dict[int, str] | None = None,
+        top_ranked: TopRankedDocumentsType,
+        query_idx_to_id: dict[int, str],
     ) -> tuple[list[list[float]], list[list[int]]]:
         doc_id_to_idx = {doc_id: i for i, doc_id in enumerate(self.idxs)}
         scores_all: list[list[float]] = []
@@ -127,15 +128,17 @@ class FaissSearchIndex:
             query_id = query_idx_to_id[query_idx]
             ranked_ids = top_ranked.get(query_id)
             if not ranked_ids:
-                logger.warning(f"No top-ranked documents for query {query_id}")
+                msg = f"No top-ranked documents for query {query_id}"
+                logger.warning(msg)
+                warnings.warn(msg)
                 scores_all.append([])
                 idxs_all.append([])
                 continue
 
             candidate_indices = [doc_id_to_idx[doc_id] for doc_id in ranked_ids]
-            d = self.index.d
+            d = self.index.d  # type: ignore[union-attr]
             candidate_embs = np.vstack(
-                [self.index.reconstruct(idx) for idx in candidate_indices]
+                [self.index.reconstruct(idx) for idx in candidate_indices]  # type: ignore[union-attr]
             )
             sub_reranking_index = self.index_type(d)
             sub_reranking_index.add(candidate_embs)