o Zh1@srddlmZddlmZmZmZddlmZddlm Z m Z m Z dZ dZ dZGdd d e eZGd d d eZd S) )Path)AnyDictList) Embeddings) BaseModel ConfigDictFieldzVs zMOpenVINOEmbeddings.__init__..require_model_export..z/openvino_model.xmlz.xmlz.binT)ris_direxists model_infoas_posixZsiblingsreplace Exception)rrrZ model_dirZhf_apir%Z model_filesZ ov_model_pathrrr!require_model_exportEs2   z9OpenVINOEmbeddings.__init__..require_model_exportZexportT) AutoTokenizerzQUnable to import transformers, please install with `pip install -U transformers`.r )NN)super__init__Zoptimum.intel.openvinor ImportErrorZhuggingface_hubrstrrboolr Zfrom_pretrainedrr Z transformersr*r )selfrrer)r* __class__rr!r,/sn # zOpenVINOEmbeddings.__init__textrcCs`t|trttt|St|dsdSt|dks#t|dtr't|Stdd|DS)z Help function to get the length for the input text. Text can be either a list of ints (which means a single text as input), or a tuple of list of ints (representing several text inputs to the model). __len__rcSsg|]}t|qSr )lenrtr r r!r"z3OpenVINOEmbeddings._text_length..) isinstancedictr7nextitervalueshasattrintsumr0r4r r r! _text_length|s  zOpenVINOEmbeddings._text_lengthT sentences batch_sizeshow_progress_barconvert_to_numpyconvert_to_tensor mean_poolingnormalize_embeddingsc sLzddl}Wnty} ztd| d} ~ wwzddlm} Wnty1} ztd| d} ~ wwzddlWntyI} ztd| d} ~ wwdtdtd tffd d } |r[d }d } ttsgtd slgd} g| fddD} fdd| D}| dt |d| dD]a}||||}j j j dd}|jrj|dddd}n j|d|ddd}j di|}|r| ||d}n |ddddf}|rjjj|ddd}|r|}|qfdd| | D|rt r nn |r|ddD| r$dS)aw Computes sentence embeddings. :param sentences: the sentences to embed. :param batch_size: the batch size used for the computation. :param show_progress_bar: Whether to output a progress bar. :param convert_to_numpy: Whether the output should be a list of numpy vectors. :param convert_to_tensor: Whether the output should be one large tensor. :param mean_pooling: Whether to pool returned vectors. :param normalize_embeddings: Whether to normalize returned vectors. :return: By default, a 2d numpy array with shape [num_inputs, output_dimension]. rNzCUnable to import numpy, please install with `pip install -U numpy`.)trangezAUnable to import tqdm, please install with `pip install -U tqdm`.zCUnable to import torch, please install with `pip install -U torch`. model_outputattention_maskrcsD|d}|d|}||dj|dddS)Nrr6g& .>)min)Z unsqueezeexpandsizefloatrBclamp)rNrOZtoken_embeddingsZinput_mask_expanded)torchr r!run_mean_poolings z3OpenVINOEmbeddings.encode..run_mean_poolingFr5Tcsg|]}| qSr )rD)rZsenr0r r!r"sz-OpenVINOEmbeddings.encode..cg|]}|qSr r ridx)rFr r!r"r:ZBatches)Zdescdisabler6pt)padding truncationreturn_tensors max_length)r^rar_r`)pdimcrYr r rZ)all_embeddingsr r!r"r:cSsg|]}|qSr )numpy)rZembr r r!r"r:r )rfr-ZtqdmrMrVrr;r.r@Zargsortr7r requestZinputsZget_partial_shapeZ is_dynamicr Z get_lengthnnZ functional normalizecpuextendstackZTensorZasarray)r0rFrGrHrIrJrKrLnpr1rMrWZinput_was_stringZlength_sorted_idxZsentences_sorted start_indexZsentences_batchlengthfeaturesZ out_features embeddingsr )rer0rFrVr!encodes         zOpenVINOEmbeddings.encodeZforbidr )extraZprotected_namespacestextscCs4ttdd|}|j|fd|ji|j}|S)Compute doc embeddings using a HuggingFace transformer model. Args: texts: The list of texts to embed. Returns: List of embeddings, one for each text. cSs |ddS)N  )r')xr r r! s z4OpenVINOEmbeddings.embed_documents..rH)listmaprrrrtolistr0rtrqr r r!embed_documentss z"OpenVINOEmbeddings.embed_documentscCs||gdS)Compute query embeddings using a HuggingFace transformer model. Args: text: The text to embed. Returns: Embeddings for the text. r)r~rCr r r! embed_querys zOpenVINOEmbeddings.embed_query model_pathcCs&|j|j||j|dS)NT)r ZhalfZsave_pretrainedr )r0rr r r! save_models   zOpenVINOEmbeddings.save_model)rEFTFFT)__name__ __module__ __qualname____doc__r r__annotations__r r.r r<rrrrr/r,rArDrrrZ model_configrrTr~rr __classcell__r r r2r!r s^    M s r csxeZdZUdZeZeed< dZeed< de ffdd Z de ed e e e fd d Z d ed e e fd dZZS)OpenVINOBgeEmbeddingsaOpenVNO BGE embedding models. Bge Example: .. code-block:: python from langchain_community.embeddings import OpenVINOBgeEmbeddings model_name = "BAAI/bge-large-en-v1.5" model_kwargs = {'device': 'CPU'} encode_kwargs = {'normalize_embeddings': True} ov = OpenVINOBgeEmbeddings( model_name_or_path=model_name, model_kwargs=model_kwargs, encode_kwargs=encode_kwargs ) query_instructionembed_instructionrc s*tjdi|d|jvrt|_dSdS)rz-zhNr )r+r,r DEFAULT_QUERY_BGE_INSTRUCTION_ZHr)r0rr2r r!r,@s  zOpenVINOBgeEmbeddings.__init__rtrcs.fdd|D}j|fij}|S)rucsg|] }j|ddqS)rvrw)rr'r8rXr r!r"Psz9OpenVINOBgeEmbeddings.embed_documents..)rrrr|r}r rXr!r~Gs z%OpenVINOBgeEmbeddings.embed_documentsr4cCs.|dd}|j|j|fi|j}|S)rrvrw)r'rrrrr|)r0r4Z embeddingr r r!rTs z!OpenVINOBgeEmbeddings.embed_query)rrrr DEFAULT_QUERY_BGE_INSTRUCTION_ENrr.rrrr,rrTr~rrr r r2r!r)s    rN)pathlibrtypingrrrZlangchain_core.embeddingsrZpydanticrrr ZDEFAULT_QUERY_INSTRUCTIONrrr rr r r r!s