from haystack.nodes import TransformersQueryClassifier, Docs2Answers from haystack.nodes import EmbeddingRetriever, FARMReader from haystack.nodes.base import BaseComponent from haystack.document_stores import InMemoryDocumentStore from markdown import markdown from annotated_text import annotation from haystack.schema import Document from typing import List, Text, Union from typing_extensions import Literal from utils.preprocessing import processingpipeline from utils.streamlitcheck import check_streamlit from haystack.pipelines import Pipeline import pandas as pd import logging try: from termcolor import colored except: pass try: import streamlit as st except ImportError: logging.info("Streamlit not installed") @st.cache(allow_output_mutation=True) def loadQueryClassifier(): """ retuns the haystack query classifier model model = shahrukhx01/bert-mini-finetune-question-detection """ query_classifier = TransformersQueryClassifier(model_name_or_path= "shahrukhx01/bert-mini-finetune-question-detection") return query_classifier class QueryCheck(BaseComponent): """ Uses Query Classifier from Haystack, process the query based on query type. Ability to determine the statements is not so good, therefore the chances statement also get modified. Ex: "List water related issues" will be identified by the model as keywords, and therefore it be processed as "what are the 'list all water related issues' related issues and discussions?". This is one shortcoming but is igonred for now, as semantic search will not get affected a lot, by this. If you want to pass keywords list and want to do batch processing use. run_batch. Example: if you want to find relevant passages for water, food security, poverty then querylist = ["water", "food security","poverty"] and then execute QueryCheck.run_batch(queries = querylist) 1. https://docs.haystack.deepset.ai/docs/query_classifier """ outgoing_edges = 1 def run(self, query:str): """ mandatory method to use the custom node. Determines the query type, if if the query is of type keyword/statement will modify it to make it more useful for sentence transoformers. Params -------- query: query/statement/keywords in form of string Return ------ output: dictionary, with key as identifier and value could be anything we need to return. In this case the output contain key = 'query'. output_1: As there is only one outgoing edge, we pass 'output_1' string """ query_classifier = loadQueryClassifier() result = query_classifier.run(query=query) if result[1] == "output_1": output = {"query":query, "query_type": 'question/statement'} else: output = {"query": "what are the {} related issues and \ discussions?".format(query), "query_type": 'statements/keyword'} logging.info(output) return output, "output_1" def run_batch(self, queries:List[str]): """ running multiple queries in one go, howeevr need the queries to be passed as list of string. Example: if you want to find relevant passages for water, food security, poverty then querylist = ["water", "food security", "poverty"] and then execute QueryCheck.run_batch(queries = querylist) Params -------- queries: queries/statements/keywords in form of string encapsulated within List Return ------ output: dictionary, with key as identifier and value could be anything we need to return. In this case the output contain key = 'queries'. output_1: As there is only one outgoing edge, we pass 'output_1' string """ query_classifier = loadQueryClassifier() query_list = [] for query in queries: result = query_classifier.run(query=query) if result[1] == "output_1": query_list.append(query) else: query_list.append("what are the {} related issues and \ discussions?".format(query)) output = {'queries':query_list} logging.info(output) return output, "output_1" @st.cache(allow_output_mutation=True) def runSemanticPreprocessingPipeline(file_path:str, file_name:str, split_by: Literal["sentence", "word"] = 'sentence', split_length:int = 2, split_overlap:int = 0, split_respect_sentence_boundary:bool = False, remove_punc:bool = False)->List[Document]: """ creates the pipeline and runs the preprocessing pipeline. Params ------------ file_name: filename, in case of streamlit application use st.session_state['filename'] file_path: filepath, in case of streamlit application use st.session_state['filepath'] split_by: document splitting strategy either as word or sentence split_length: when synthetically creating the paragrpahs from document, it defines the length of paragraph. split_overlap: Number of words or sentences that overlap when creating the paragraphs. This is done as one sentence or 'some words' make sense when read in together with others. Therefore the overlap is used. split_respect_sentence_boundary: Used when using 'word' strategy for splititng of text. remove_punc: to remove all Punctuation including ',' and '.' or not Return -------------- List[Document]: When preprocessing pipeline is run, the output dictionary has four objects. For the Haysatck implementation of semantic search we, need to use the List of Haystack Document, which can be fetched by key = 'documents' on output. """ semantic_processing_pipeline = processingpipeline() output_semantic_pre = semantic_processing_pipeline.run(file_paths = file_path, params= {"FileConverter": {"file_path": file_path, \ "file_name": file_name}, "UdfPreProcessor": {"remove_punc": remove_punc, \ "split_by": split_by, \ "split_length":split_length,\ "split_overlap": split_overlap, "split_respect_sentence_boundary":split_respect_sentence_boundary}}) return output_semantic_pre @st.cache(hash_funcs={"builtins.SwigPyObject": lambda _: None}, allow_output_mutation=True) def loadRetriever(embedding_model:Text=None, embedding_model_format:Text = None, embedding_layer:int = None, retriever_top_k:int = 10, max_seq_len:int=512, document_store:InMemoryDocumentStore=None): """ Returns the Retriever model based on params provided. 1. https://docs.haystack.deepset.ai/docs/retriever#embedding-retrieval-recommended 2. https://www.sbert.net/examples/applications/semantic-search/README.html 3. https://github.com/deepset-ai/haystack/blob/main/haystack/nodes/retriever/dense.py Params --------- embedding_model: Name of the model to be used for embedding. Check the links provided in documentation embedding_model_format: check the github link of Haystack provided in documentation embedding_layer: check the github link of Haystack provided in documentation retriever_top_k: Number of Top results to be returned by retriever max_seq_len: everymodel has max seq len it can handle, check in model card. Needed to hanlde the edge cases. document_store: InMemoryDocumentStore, write haystack Document list to DocumentStore and pass the same to function call. Can be done using createDocumentStore from utils. Return ------- retriever: embedding model """ logging.info("loading retriever") if document_store is None: logging.warning("Retriever initialization requires the DocumentStore") return retriever = EmbeddingRetriever( embedding_model=embedding_model,top_k = retriever_top_k, document_store = document_store, emb_extraction_layer=embedding_layer, scale_score =True, model_format=embedding_model_format, use_gpu = True, max_seq_len = max_seq_len ) if check_streamlit: st.session_state['retriever'] = retriever return retriever @st.cache(hash_funcs={"builtins.SwigPyObject": lambda _: None}, allow_output_mutation=True) def createDocumentStore(documents:List[Document], similarity:str = 'dot_product', embedding_dim:int = 768): """ Creates the InMemory Document Store from haystack list of Documents. It is mandatory component for Retriever to work in Haystack frame work. Params ------- documents: List of haystack document. If using the preprocessing pipeline, can be fetched key = 'documents; on output of preprocessing pipeline. similarity: scoring function, can be either 'cosine' or 'dot_product' embedding_dim: Document store has default value of embedding size = 768, and update_embeddings method of Docstore cannot infer the embedding size of retiever automatically, therefore set this value as per the model card. Return ------- document_store: InMemory Document Store object type. """ document_store = InMemoryDocumentStore(similarity = similarity, embedding_dim = embedding_dim ) document_store.write_documents(documents) return document_store @st.cache(hash_funcs={"builtins.SwigPyObject": lambda _: None}, allow_output_mutation=True) def semanticSearchPipeline(documents:List[Document], embedding_model:Text = None, embedding_model_format:Text = None,embedding_layer:int = None, embedding_dim:int = 768,retriever_top_k:int = 10, reader_model:str = None, reader_top_k:int = 10, max_seq_len:int =512,useQueryCheck = True, ): """ creates the semantic search pipeline and document Store object from the list of haystack documents. The top_k for the Reader and Retirever are kept same, so that all the results returned by Retriever are used, however the context is extracted by Reader for each retrieved result. The querycheck is added as node to process the query. This pipeline is suited for keyword search, and to some extent extractive QA purpose. The purpose of Reader is strictly to highlight the context for retrieved result and not for QA, however as stated it can work for QA too in limited sense. There are 4 variants of pipeline it can return 1.QueryCheck > Retriever > Reader 2.Retriever > Reader 3.QueryCheck > Retriever > Docs2Answers : If reader is None, then Doc2answer is used to keep the output of pipeline structurally same. 4.Retriever > Docs2Answers Links 1. https://docs.haystack.deepset.ai/docs/retriever#embedding-retrieval-recommended 2. https://www.sbert.net/examples/applications/semantic-search/README.html 3. https://github.com/deepset-ai/haystack/blob/main/haystack/nodes/retriever/dense.py 4. https://docs.haystack.deepset.ai/docs/reader Params ---------- documents: list of Haystack Documents, returned by preprocessig pipeline. embedding_model: Name of the model to be used for embedding. Check the links provided in documentation embedding_model_format: check the github link of Haystack provided in documentation embedding_layer: check the github link of Haystack provided in documentation embedding_dim: Document store has default value of embedding size = 768, and update_embeddings method of Docstore cannot infer the embedding size of retiever automatically, therefore set this value as per the model card. retriever_top_k: Number of Top results to be returned by retriever reader_model: Name of the model to be used for Reader node in hasyatck Pipeline. Check the links provided in documentation reader_top_k: Reader will use retrieved results to further find better matches. As purpose here is to use reader to extract context, the value is same as retriever_top_k. max_seq_len:everymodel has max seq len it can handle, check in model card. Needed to hanlde the edge cases useQueryCheck: Whether to use the querycheck which modifies the query or not. Return --------- semanticsearch_pipeline: Haystack Pipeline object, with all the necessary nodes [QueryCheck, Retriever, Reader/Docs2Answer]. If reader is None, then Doc2answer is used to keep the output of pipeline structurally same. document_store: As retriever can work only with Haystack Document Store, the list of document returned by preprocessing pipeline are fed into to get InMemmoryDocumentStore object type, with retriever updating the embeddings of each paragraph in document store. """ document_store = createDocumentStore(documents=documents, embedding_dim=embedding_dim) retriever = loadRetriever(embedding_model = embedding_model, embedding_model_format=embedding_model_format, embedding_layer=embedding_layer, retriever_top_k= retriever_top_k, document_store = document_store, max_seq_len=max_seq_len) document_store.update_embeddings(retriever) semantic_search_pipeline = Pipeline() if useQueryCheck and reader_model: querycheck = QueryCheck() reader = FARMReader(model_name_or_path=reader_model, top_k = reader_top_k, use_gpu=True) semantic_search_pipeline.add_node(component = querycheck, name = "QueryCheck",inputs = ["Query"]) semantic_search_pipeline.add_node(component = retriever, name = "EmbeddingRetriever",inputs = ["QueryCheck.output_1"]) semantic_search_pipeline.add_node(component = reader, name = "FARMReader", inputs= ["EmbeddingRetriever"]) elif reader_model : reader = FARMReader(model_name_or_path=reader_model, top_k = reader_top_k, use_gpu=True) semantic_search_pipeline.add_node(component = retriever, name = "EmbeddingRetriever",inputs = ["Query"]) semantic_search_pipeline.add_node(component = reader, name = "FARMReader",inputs= ["EmbeddingRetriever"]) elif useQueryCheck and not reader_model: querycheck = QueryCheck() docs2answers = Docs2Answers() semantic_search_pipeline.add_node(component = querycheck, name = "QueryCheck",inputs = ["Query"]) semantic_search_pipeline.add_node(component = retriever, name = "EmbeddingRetriever",inputs = ["QueryCheck.output_1"]) semantic_search_pipeline.add_node(component = docs2answers, name = "Docs2Answers",inputs= ["EmbeddingRetriever"]) elif not useQueryCheck and not reader_model: docs2answers = Docs2Answers() semantic_search_pipeline.add_node(component = retriever, name = "EmbeddingRetriever",inputs = ["Query"]) semantic_search_pipeline.add_node(component = docs2answers, name = "Docs2Answers",inputs= ["EmbeddingRetriever"]) logging.info(semantic_search_pipeline.components) return semantic_search_pipeline, document_store def runSemanticPipeline(pipeline:Pipeline, queries:Union[list,str])->dict: """ will use the haystack run or run_batch based on if single query is passed as string or multiple queries as List[str] Params ------- pipeline: haystack pipeline, this is same as returned by semanticSearchPipeline from utils.semanticsearch queries: Either a single query or list of queries. Return ------- results: Dict containing answers and documents as key and their respective values """ if type(queries) == list: results = pipeline.run_batch(queries=queries) elif type(queries) == str: results = pipeline.run(query=queries) else: logging.info("Please check the input type for the queries") return return results def process_query_output(results:dict)->pd.DataFrame: """ Returns the dataframe with necessary information like including ['query','answer','answer_offset','context_offset','context','content', 'reader_score','retriever_score','id',]. This is designed for output given by semantic search pipeline with single query and final node as reader. The output of pipeline having Docs2Answers as final node or multiple queries need to be handled separately. In these other cases, use process_semantic_output from utils.semantic_search which uses this function internally to make one combined dataframe. Params --------- results: this dictionary should have key,values with keys = [query,answers,documents], however answers is optional. in case of [Doc2Answers as final node], process_semantic_output doesnt return answers thereby setting all values contained in answers to 'None' Return -------- df: dataframe with all the columns mentioned in function description. """ query_text = results['query'] if 'answers' in results.keys(): answer_dict = {} for answer in results['answers']: answer_dict[answer.document_id] = answer.to_dict() else: answer_dict = {} docs = results['documents'] df = pd.DataFrame(columns=['query','answer','answer_offset','context_offset', 'context','content','reader_score','retriever_score', 'id']) for doc in docs: row_list = {} row_list['query'] = query_text row_list['retriever_score'] = doc.score row_list['id'] = doc.id row_list['content'] = doc.content if doc.id in answer_dict.keys(): row_list['answer'] = answer_dict[doc.id]['answer'] row_list['context'] = answer_dict[doc.id]['context'] row_list['reader_score'] = answer_dict[doc.id]['score'] answer_offset = answer_dict[doc.id]['offsets_in_document'][0] row_list['answer_offset'] = [answer_offset['start'],answer_offset['end']] start_idx = doc.content.find(row_list['context']) end_idx = start_idx + len(row_list['context']) row_list['context_offset'] = [start_idx, end_idx] else: row_list['answer'] = None row_list['context'] = None row_list['reader_score'] = None row_list['answer_offset'] = None row_list['context_offset'] = None df_dictionary = pd.DataFrame([row_list]) df = pd.concat([df, df_dictionary], ignore_index=True) return df def process_semantic_output(results): """ Returns the dataframe with necessary information like including ['query','answer','answer_offset','context_offset','context','content', 'reader_score','retriever_score','id',]. Distingushes if its single query or multi queries by reading the pipeline output dictionary keys. Uses the process_query_output to get the dataframe for each query and create one concataneted dataframe. In case f Docs2Answers as final node, deletes the answers part. See documentations of process_query_output. Params --------- results: raw output of runSemanticPipeline. Return -------- df: dataframe with all the columns mentioned in function description. """ output = {} if 'query' in results.keys(): output['query'] = results['query'] output['documents'] = results['documents'] if results['node_id'] == 'Docs2Answers': pass else: output['answers'] = results['answers'] df = process_query_output(output) return df if 'queries' in results.keys(): df = pd.DataFrame(columns=['query','answer','answer_offset', 'context_offset','context','content', 'reader_score','retriever_score','id']) for query,answers,documents in zip(results['queries'], results['answers'],results['documents']): output = {} output['query'] = query output['documents'] = documents if results['node_id'] == 'Docs2Answers': pass else: output['answers'] = answers temp = process_query_output(output) df = pd.concat([df, temp], ignore_index=True) return df def semanticsearchAnnotator(matches:List[List[int]], document:Text): """ Annotates the text in the document defined by list of [start index, end index] Example: "How are you today", if document type is text, matches = [[0,3]] will give answer = "How", however in case we used the spacy matcher then the matches = [[0,3]] will give answer = "How are you". However if spacy is used to find "How" then the matches = [[0,1]] for the string defined above. """ start = 0 annotated_text = "" for match in matches: start_idx = match[0] end_idx = match[1] if check_streamlit(): annotated_text = (annotated_text + document[start:start_idx] + str(annotation(body=document[start_idx:end_idx], label="Context", background="#964448", color='#ffffff'))) else: annotated_text = (annotated_text + document[start:start_idx] + colored(document[start_idx:end_idx], "green", attrs = ['bold'])) start = end_idx annotated_text = annotated_text + document[end_idx:] if check_streamlit(): st.write( markdown(annotated_text), unsafe_allow_html=True, ) else: print(annotated_text) def semantic_keywordsearch(query:Text,documents:List[Document], embedding_model:Text, embedding_model_format:Text, embedding_layer:int, reader_model:str, retriever_top_k:int = 10, reader_top_k:int = 10, return_results:bool = False, embedding_dim:int = 768, max_seq_len:int = 512, sort_by:Literal["retriever", "reader"] = 'retriever'): """ Performs the Semantic search on the List of haystack documents which is returned by preprocessing Pipeline. Params ------- query: Keywords that need to be searche in documents. documents: List fo Haystack documents returned by preprocessing pipeline. """ semanticsearch_pipeline, doc_store = semanticSearchPipeline(documents = documents, embedding_model= embedding_model, embedding_layer= embedding_layer, embedding_model_format= embedding_model_format, reader_model= reader_model, retriever_top_k= retriever_top_k, reader_top_k= reader_top_k, embedding_dim=embedding_dim, max_seq_len=max_seq_len) raw_output = runSemanticPipeline(semanticsearch_pipeline,query) results_df = process_semantic_output(raw_output) if sort_by == 'retriever': results_df = results_df.sort_values(by=['retriever_score'], ascending=False) else: results_df = results_df.sort_values(by=['reader_score'], ascending=False) if return_results: return results_df else: if check_streamlit: st.markdown("##### Top few semantic search results #####") else: print("Top few semantic search results") for i in range(len(results_df)): if check_streamlit: st.write("Result {}".format(i+1)) else: print("Result {}".format(i+1)) semanticsearchAnnotator(results_df.loc[i]['context_offset'], results_df.loc[i]['content'] )