+#h(dZddlmZmZmZddlZddlZddlmZ ddl Z e e dddl Z e e jjddk(Z ddlZ e j$j&e j$j(e j$j*fZGddej.Zy#e $rdZ YwxYw#e$rdZYrwxYw#e$re j&e j*fZYOwxYw) z< Retriever model for Pinecone Author: Dhar Rawal (@drawal1) )ListOptionalUnionN)dotdictzcThe pinecone library is required to use PineconeRM. Install it with `pip install dspy-ai[pinecone]`TcneZdZdZ ddedeedeedeedeedeedeed effd Z dd ed eed eedeedeedejf dZ dZ e je jeddeedeeefdZdeeeefdej*fdZxZS) PineconeRMa  A retrieval module that uses Pinecone to return the top passages for a given query. Assumes that the Pinecone index has been created and populated with the following metadata: - text: The text of the passage Args: pinecone_index_name (str): The name of the Pinecone index to query against. pinecone_api_key (str, optional): The Pinecone API key. Defaults to None. pinecone_env (str, optional): The Pinecone environment. Defaults to None. local_embed_model (str, optional): The local embedding model to use. A popular default is "sentence-transformers/all-mpnet-base-v2". openai_embed_model (str, optional): The OpenAI embedding model to use. Defaults to "text-embedding-ada-002". openai_api_key (str, optional): The API key for OpenAI. Defaults to None. openai_org (str, optional): The organization for OpenAI. Defaults to None. k (int, optional): The number of top passages to retrieve. Defaults to 3. Returns: dspy.Prediction: An object containing the retrieved passages. Examples: Below is a code snippet that shows how to use this as the default retriver: ```python llm = dspy.OpenAI(model="gpt-3.5-turbo") retriever_model = PineconeRM(openai.api_key) dspy.settings.configure(lm=llm, rm=retriever_model) ``` Below is a code snippet that shows how to use this in the forward() function of a module ```python self.retrieve = PineconeRM(k=num_passages) ``` pinecone_index_namepinecone_api_key pinecone_envlocal_embed_modelopenai_embed_modelopenai_api_key openai_orgkc 2| ddl} ddlm} m} | j ||_| j ||_d|_ | j| jjrdn'| jjjrdnd|_ n6|)||_d|_ |r |t _|r|t _n t'd |j)||||_t, |]| y#t$r} t d| d} ~ wwxYw) Nr) AutoModel AutoTokenizerzeYou need to install Hugging Face transformers library to use a local embedding model with PineconeRM.Tzcuda:0mpscpuFz@Either local_embed_model or openai_embed_model must be provided.)r)torch transformersrr ImportErrorModuleNotFoundErrorfrom_pretrained_local_embed_model_local_tokenizeruse_local_modeldevicecuda is_availablebackendsr_openai_embed_modelopenaiapi_key organization ValueError_init_pinecone_pinecone_indexsuper__init__)selfr r r r r rrrrrrexc __class__s V/var/www/html/sandstorm/venv/lib/python3.12/site-packages/dspy/retrieve/pinecone_rm.pyr*zPineconeRM.__init__Fs  ( A '0&?&?@Q&RD #$1$A$ABS$TD !#'D ,,!JJ335++88:DK  +'9D $#(D !/&0#R  $22 !1<  1= )w s C<< D DD index_namer$ environment dimensiondistance_metricreturnci}|r||d<|r||d<tjdi|tj}||vr'| | tdtj|||tj |S)anInitialize pinecone and return the loaded index. Args: index_name (str): The name of the index to load. If the index is not does not exist, it will be created. api_key (str, optional): The Pinecone API key, defaults to env var PINECONE_API_KEY if not provided. environment (str, optional): The environment (ie. `us-west1-gcp` or `gcp-starter`. Defaults to env PINECONE_ENVIRONMENT. Raises: ValueError: If api_key or environment is not provided and not set as an environment variable. Returns: pinecone.Index: The loaded index. r$r0zWdimension and distance_metric must be provided since the index provided does not exist.)namer1metric)pineconeinit list_indexesr& create_indexIndex)r+r/r$r0r1r2kwargsactive_indexess r.r'zPineconeRM._init_pineconeus.  'F9  $/F= ! !..0 ^ + _%< m  ! !#&  ~~j))c8 ddl}|d}|jdj |j j }|j||zd|j|jddz S#t$r}td|d}~wwxYw)NrIYou need to install torch to use a local embedding model with PineconeRM.g& .>)min) rrr unsqueezeexpandsizefloatsumclamp)r+ model_outputattention_maskrr,token_embeddingsinput_mask_expandeds r. _mean_poolingzPineconeRM._mean_poolings    (?,66r:AABRBWBWBYZ``byy),??CekkReRiRijkRlrvkFwww %[  sA?? B BB)max_timequeriesc8 ddl}|jstr,t j j||j}n9t jj||jj}|dDcgc]}|d c}S|j|ddd j|j}|j5|jdi|j|j}ddd|j!|d }|j"j$j'|d d }|j)j+j-S#t$r}td|d}~wwxYwcc}w#1swYxYw)zReturn query vector after creating embedding using OpenAI Args: queries (list): List of query strings to embed. Returns: List[List[float]]: List of embeddings corresponding to each query. rNrA)inputmodeldata embeddingTpt)padding truncationreturn_tensorsrLrC)pdimr7)rrrr OPENAI_LEGACYr# Embeddingcreater" embeddings model_dumprtorno_gradrrOnn functional normalizernumpytolist) r+rRrr,rW encoded_inputrKrbnormalized_embeddingss r._get_embeddingszPineconeRM._get_embeddingss"   ##",,33!)A)A4 #--44!)A)A5*,=Ffz$PineconeRM.forward..squuXs7Kr?)keyreversemetadatatext long_text)passagesruscorec |dS)NrCr7rws r.ryz$PineconeRM.forward..s !A$r?N) isinstancestrrmlenr(queryrsortedrdspy Predictionrvitems)r+rnrRqrb results_dictsorted_resultsresultrpassagepassage_scoresrWsorted_passages_s r.forwardzPineconeRM.forwards*C0 !  &+1++))'2  w<1 //551 TVVd6L $Y'-KUYNBPPvz*62PHPX N'g!5 NOPH??H5 5# I//55!d6L'y1 "&&vj'9&'A3GWo&vj1&9:   !  " DFF[j(kZWVW+w1G)H(kllC,Q N&)lsF F F" F-F )NNNztext-embedding-ada-002NNr)NNNN)__name__ __module__ __qualname____doc__rrintr*r8r<r'rObackoff on_exceptionexpoERRORSrrHrmrrrr __classcell__)r-s@r.rr$siH+/&*+/,D(,$(- -#3--sm - $C= - %SM -! -SM- -d"&%)#')- +*+*#+*c] +* C= +* "# +* +*Zx W  %<c%< d5k %<  %rs )(   m 22156!;M6ll))6<<+O+OQWQ]Q]QfQf gFlmlm+HM 6##V__ 5F6s5B-"B:AC-B76B7:CCC'&C'