import os import whisper from io import BytesIO # BytesIO is a class in the io module that implements an in-memory file-like object. import base64 import boto3 # AWS Polly from pydub import AudioSegment # AudioSegment is a class in the pydub module that can be used to manipulate audio files. from pydub.playback import play # play is a function in the pydub.playback module that can be used to play audio files. import logging import numpy as np import openai from langchain import OpenAI from langchain.chains import RetrievalQA # RetrievalQA is a class in the langchain.chains module that can be used to build a retrieval-based question answering system. from langchain.vectorstores import Chroma # Chroma is a class in the langchain.vectorstores module that can be used to store vectors. from langchain.document_loaders import DirectoryLoader # from langchain.embeddings.openai import OpenAIEmbeddings # OpenAIGPTEmbeddings from langchain.embeddings import HuggingFaceEmbeddings from langchain.text_splitter import CharacterTextSplitter # CharacterTextSplitter is a class in the langchain.text_splitter module that can be used to split text into chunks. #import streamlit as st from langchain.indexes import VectorstoreIndexCreator #导入向量存储索引创建器 #from langchain.vectorstores import DocArrayInMemorySearch #向量存储 from tenacity import ( retry, stop_after_attempt, wait_random_exponential, ) # FUNCTIONS # get embeddings @retry(wait=wait_random_exponential(min=21, max=60), stop=stop_after_attempt(100)) #@st.cache_data def embedding_from_string(input: str, model: str) -> list: response = openai.Embedding.create(input=input, model=model) embedding = response["data"][0]["embedding"] return embedding OPENAI_API_KEY = os.getenv('OPENAI_API_KEY') AWS_ACCESS_KEY_ID = os.getenv('AWS_ACCESS_KEY_ID') AWS_SECRET_ACCESS_KEY = os.getenv('AWS_SECRET_ACCESS_KEY') AWS_REGION_NAME = 'ap-south-1' logging.basicConfig(level="INFO", filename='conversations.log', filemode='a', format='%(asctime)s %(message)s', datefmt='%H:%M:%S') def buzz_user(): input_prompt = AudioSegment.from_mp3('assets/timeout_audio.mp3') play(input_prompt) def initialize_knowledge_base(): loader = DirectoryLoader('profiles', glob='**/*.txt') #文件夹加载器 profiles文件夹下的所有txt文件 docs = loader.load() #index = VectorstoreIndexCreator( # vectorstore_cls=DocArrayInMemorySearch #).from_loaders([loader]) char_text_splitter = CharacterTextSplitter(chunk_size=1000, chunk_overlap=0) #文本分割器 chunk_size=1000, chunk_overlap=0 doc_texts = char_text_splitter.split_documents(docs) #文档分割器,作用是将文档分割成小块 # Embed each chunk of text #embeddings = [] #openAI_embeddings = OpenAIEmbeddings() #for doc in doc_texts: # text = str(doc) #embedding = openAI_embeddings.embed_documents(text) #embeddings.append(embedding) # embedding = embedding_from_string(text, "text-embedding-ada-002") # embeddings.append(embedding) #vStore = np.concatenate(embeddings, axis=0) embedding = HuggingFaceEmbeddings(model_name='shibing624/text2vec-base-chinese') #openAI_embeddings = OpenAIEmbeddings() vStore = Chroma.from_documents(doc_texts, embedding) #Chroma是一个类,用于存储向量,from_documents是一个方法,用于从文档中创建向量存储器,openAI_embeddings是一个类,用于将文本转换为向量 conv_model = RetrievalQA.from_chain_type( llm=OpenAI(model_name="gpt-3.5-turbo-16k"), chain_type="stuff", retriever=vStore.as_retriever( search_kwargs={"k": 1} ) ) voice_model = whisper.load_model("tiny") #加载模型 tiny模型 tiny模型是一个小型的语音识别模型,它的大小只有 50MB 左右,但是它的准确率却非常高,可以达到 95% 以上。 return conv_model, voice_model def text_to_speech_gen(answer): #文字转语音 polly = boto3.client('polly', aws_access_key_id=AWS_ACCESS_KEY_ID, aws_secret_access_key=AWS_SECRET_ACCESS_KEY, region_name=AWS_REGION_NAME) response = polly.synthesize_speech( Text=answer, #VoiceId='Matthew', VoiceId='Zhiyu', OutputFormat='mp3', #Engine = "neural" Engine = "standard") audio_stream = response['AudioStream'].read() audio_html = audio_to_html(audio_stream) return audio_html def audio_to_html(audio_bytes): #音频转html audio_io = BytesIO(audio_bytes) audio_io.seek(0) audio_base64 = base64.b64encode(audio_io.read()).decode("utf-8") audio_html = f'' return audio_html def get_chat_history(user_message, history): return "", history + [[user_message, None]]