'XhHddlZddlmZddlmZddlmZddlmZm Z m Z m Z m Z m Z ddlmZddlmZddlmZdd lmZdd lmZdd lmZmZdd lmZdd lmZddlm Z ddl!m"Z"m#Z#ddl$m%Z%Gdde#Z&GddZ'y)N) defaultdict)deepcopy)get_all_start_methods)OptionalUnionIterableAnyTypeget_args) BaseModel)models)INFERENCE_OBJECT_TYPES)InspectorEmbed)Embedder) NumericVectorNumericVectorStruct)ModelSchemaParser) FieldPath) FastEmbedMisc)ParallelWorkerPoolWorker) iter_batchcjeZdZdedefdZedededdfdZdee eefdee eeffdZ y) ModelEmbedderWorker batch_sizekwargsc 2tdi||_||_y)N) ModelEmbeddermodel_embedderr)selfrrs L/RAG/venv/lib/python3.12/site-packages/qdrant_client/embed/model_embedder.py__init__zModelEmbedderWorker.__init__s+5f5$returnc |dd|d|S)N)threadsrrr)clsrrs r"startzModelEmbedderWorker.starts>1>v>>r$itemsc#K|D]9\}}|t|jj||jf;yw)Ninference_batch_size)listr embed_models_batchr)r!r+idxbatchs r"processzModelEmbedderWorker.processsKJC''::DOO;  sAAN) __name__ __module__ __qualname__intr r# classmethodr*rtupler3rr$r"rrsl%3%#%?s?c?6K?? XeCHo6 8E#s(O;T r$rc<eZdZdZddeedefdZ ddee e e fde de d e e fd Z dde eeee fe fde d ee d e eeee fe ffd Z ddeeeee fe fde d e d e e fdZ ddeeee fe fdeeede de d ee d eeee feeefe eff dZdej.d dfdZ ddej.de d e d efdZdde d e d dfdZded efdZedej.d ej.fdZed e e!fdZ"y) r@Nparserrc bi|_i|_t||_t di||_y)N)r<r)_batch_accumulator_embed_storager_embed_inspectorrembedder)r!r<rs r"r#zModelEmbedder.__init__-s.KM>@ .f = *6* r$ raw_modelsis_queryrr%c#Kt|tr|g}t||D]}|j|||Ed{ y7w)a2Embed raw data fields in models and return models with vectors If any of model fields required inference, a deepcopy of a model with computed embeddings is returned, otherwise returns original models. Args: raw_models: Iterable[BaseModel] - models which can contain fields with raw data is_query: bool - flag to determine which embed method to use. Defaults to False. batch_size: int - batch size for inference Returns: list[BaseModel]: models with embedded fields r-N) isinstancer rr0)r!rBrCrraw_models_batchs r" embed_modelszModelEmbedder.embed_models3sU" j) ,$J *:z B .. (/  !C s:AAAparallelc#Kd}t|trt||krd}||dk(s|r-t||D]}|j ||Ed{yd}t||}|dk(rt j }dtvrdnd }|Jt||j||j } | j|| D] }|Ed{y77 w) a1Embed raw data fields in models and return models with vectors Requires every input sequences element to contain raw data fields to inference. Does not accept ready vectors. Args: raw_models: Iterable[BaseModel] - models which contain fields with raw data to inference batch_size: int - batch size for inference parallel: int - number of parallel processes to use. Defaults to None. Returns: Iterable[Union[dict[str, BaseModel], BaseModel]]: models with embedded fields FTNr'r-)sizer forkserverspawn) num_workersworker start_methodmax_internal_batch_size)r) rEr/lenrr0os cpu_countrr_get_worker_classMAX_INTERNAL_BATCH_SIZE ordered_map) r!rBrrHis_smallr2multiprocessing_batch_sizeraw_models_batchesrOpools r"embed_models_strictz!ModelEmbedder.embed_models_strictKs& j$ ':+  x1}#J ;225z2ZZZ<*+ &!+J=W!X 1}<<>+7;P;R+Rz3ModelEmbedder.embed_models_batch..s9",I ##%!&)= $ ",s")r`r>)r!rBrCr.rbs` `` r"r0z ModelEmbedder.embed_models_batch|s`"$I    H4  P$&&! ! !",   " s!-AAAA AAmodelpathsr]c t|ttr1|r|j|n|Jd|j |||S||s t |n|}t|t rJ|jD]5\}}|r|j||d|j|||d|||<7|S||n|jj|}|D]}t|ts|gn|} | D]} t| |jd} | |jr!|j| |j|||Jt| t} | r| n| g} |s_|Jd| D cgc]} |j | ||}} | rt| |j|t| |j|d| D]} |j| |Scc} w) abEmbed model's fields requiring inference Args: model: Qdrant http model containing fields to embed paths: Path to fields to embed. E.g. [FieldPath(current="recommend", tail=[FieldPath(current="negative", tail=None)])] is_query: Flag to determine which embed method to use. Defaults to False. accumulating: Flag to determine if we are accumulating models for batch embedding. Defaults to False. inference_batch_size: Optional[int] - batch size for inference Returns: A deepcopy of the method with embedded fields Nz3inference_batch_size should be passed for inferencerCr.T)r]Fr_r)rEr r _accumulate_drain_accumulatorrdictr+r`r@inspectr/getattrcurrenttailsetattr)r!rdrerCr]r.keyvaluepath list_modelitem current_modelwas_listdata embeddingss r"r`zModelEmbedder._process_models$* eX&<= >  ')4IHI4..%)=/ =+7HUOUE eT "#kkm U''u4'H!%!4!4!)%*-A "5"E#J ,L*0E0E0M0Me0TD(25$(?%UJ" 'dllD A  (99''% !)%1-A ( *->H5=MM?M'0<QPQ< )6 &)6!33 $xNb4)6 #& $#D$,, C#D$,, 1 F$1D ,,T2%2?#F &s!Grwct|tr'|jD]}|j|yt|tr3|D].}t|t t sy|j|0t|t t sy|j|}|j|jvrg|j|j<|j|jj|y)aHAdd data to batch accumulator Args: data: models.VectorStruct - any vector struct data, if inference object types instances in `data` - add them to the accumulator, otherwise - do nothing. `InferenceObject` instances are converted to proper types. Returns: None N) rErjvaluesrhr/r r_resolve_inference_objectrdr>append)r!rwrqs r"rhzModelEmbedder._accumulates dD !  '' dD !!%2H)IJ  ' $)? @A--d3 ::T44 424D # #DJJ /  +2248r$c,t|tr1|jD]\}}|j|||||<|St|trIt |D]9\}}t|t ts|cS|j|||||<;|St|t ts|S|jr&|jj|jds|j|||j|jS)atDrain accumulator and replaces inference objects with computed embeddings It is assumed objects are traversed in the same order as they were added to the accumulator Args: data: models.VectorStruct - any vector struct data, if inference object types instances in `data` - replace them with computed embeddings. If embeddings haven't yet been computed - compute them and then replace inference objects. inference_batch_size: int - batch size for inference Returns: NumericVectorStruct: data with replaced inference objects rgN) rErjr+rir/ enumerater rr?getrd_embed_accumulator _next_embed)r!rwrCr.rprqis r"riz ModelEmbedder._drain_accumulators dD !"jjl U 33HCW4S +K dD !%dO5!%2H)IJK11HCW2Q ,K (12 K""$*=*=*A*A$**d*S  # #XL` # a ++r$c &dttdtdtdttffd }j D]}t tj|tj|tj|tj|tj|frst|dj jD]\}}||||j|<j j!y) aEmbed all accumulated objects for all models Args: is_query: bool - flag to determine which embed method to use. Defaults to False. inference_batch_size: int - batch size for inference Returns: None objects model_namerr%c g}g}g}tt}t|D]\}}t|tj } tt ||D]_\} \} } | |jk(s| | k(s|| j||| j| r |jn |j|g|t|<|j|j|j| |j| r |jn |jgg} tt ||D]X\}\} } | jjj|| r||nd| s||nd| xsi|Dcgc]}|c}Zt| }ggt|z}|j!D]}|D]}t#|||<|Scc}w)z Assemble batches by options and data type based groups, embeds and return embeddings in the original order N)rtextsimagesrCoptionsr)rr/r~rEr Documentziprr|textimagerQextendrAembediterrznext)rrrunique_optionsunique_options_is_textbatches group_indicesrobjis_textjroptions_is_textrx embeddingiter_embeddingsordered_embeddingsindicesindexrCr!s r"rz/ModelEmbedder._embed_accumulator..embedIs 46N13 "!#G2=d2CM#G,3$S&//:5>(>?61A1#++-'_2L%a(//2 ))g#((399M 6;<M#n"56"))#++6*11':NNCHHSYY#GH- J)23~G]3^)_%%GW!!*.)<)<'107'!*T5<71:$%-$+Mr'1 *=* *I"*  *`#:.O=?D3w<anyris_supported_text_modelis_supported_sparse_model(is_supported_late_interaction_text_modelis_supported_image_model.is_supported_late_interaction_multimodal_model ValueErrorr+r?clear)r!rCr.rrdrws`` r"rz ModelEmbedder._embed_accumulator?s0 &010 &?B0 &PS0 & - 0 &d,,E!99%@!;;EB!JJ5Q!::5A!PPQVW !E7*H!IJJ- 2288:KE4).;O*D   &; %%'r$rc>|j|jdS)zGet next computed embedding from embedded batch Args: model_name: str - retrieve embedding from the storage by this model name Returns: NumericVector: computed embedding r)r?pop)r!rs r"rzModelEmbedder._next_embeds "":.22155r$ct|tjs|S|j}|j}|j }t tj|tj|tj|frtj|||Stj|rtj|||Stj|rt|dt|d)aResolve inference object into a model Args: data: models.VectorStruct - data to resolve, if it's an inference object, convert it to a proper type, otherwise - keep unchanged Returns: models.VectorStruct: resolved data )rdrr)rdrrz- does not support `InferenceObject` interfacer)rEr InferenceObjectrdobjectrrrrrrrrImagerr)rwrrqrs r"r{z'ModelEmbedder._resolve_inference_objects$ 6 67KZZ  ,, 55jA77 CFFzR  ??%Q Q  1 1* =<?@"  )  H,0".2 XT#y.)945XY(X X  X 'sm X tCN#T#}*<%=y-W XXt 3 3>VW(,''(,37(,OR(, (,TL(4L(sL([_L(\ 6c 6m 6H(;(;H@S@SHH@#$':";##r$r)(rR collectionsrcopyrmultiprocessingrtypingrrrr r r pydanticr qdrant_client.httpr qdrant_client.embed.commonr#qdrant_client.embed.embed_inspectorrqdrant_client.embed.embedderrqdrant_client.embed.modelsrr!qdrant_client.embed.schema_parserrqdrant_client.embed.utilsrqdrant_client.fastembed_commonr qdrant_client.parallel_processorrrqdrant_client.uploader.uploaderrrrrr$r"rsP #1AA%=>1I?/8G6&*Q#Q#r$