ConversationalRetrievalChain 检索器

Question

在下面的代码中，我在 ConversationalRetrievalChain 检索器中遇到了一个大问题，当尝试增加“k”返回的文档数量时，我已经超出了最大令牌数量，问题是默认情况下它只返回 4，而我的csv数据有1000行

`
从 googletrans 导入翻译器

import re

from deep_translator import GoogleTranslator

from langchain.embeddings import HuggingFaceEmbeddings

import os

from transformers import AutoConfig

from langchain.text_splitter import RecursiveCharacterTextSplitter

from langchain.vectorstores import FAISS

from langchain.document_loaders.csv_loader import CSVLoader

from langchain.llms import CTransformers

from langchain.chains import ConversationalRetrievalChain

import pandas as pd

from transformers import AutoModelForCausalLM

from langchain.memory import ConversationBufferMemory

from langchain_chroma import Chroma

from langchain_community.document_loaders import TextLoader

from langchain_community.embeddings.sentence_transformer import SentenceTransformerEmbeddings

from langchain_text_splitters import CharacterTextSplitter



def translate_to_portuguese(text):

    translator = GoogleTranslator(source="pt", target="en")

    translated_segments = []

    segments = re.split(r'(".*?")', text)  # Split text into segments based on quotation marks

    for segment in segments:

        if segment.startswith('"') and segment.endswith('"'):

            translated_segments.append(segment)  # If segment is within quotation marks, keep it unchanged

        else:

            translated_text = translator.translate(segment)

            if translated_text is not None:

                translated_segments.append(translated_text)

            else:

                translated_segments.append(segment)

    translated_text = ' '.join(translated_segments)

    return translated_text



def main():

    # Load the document and split it into chunks

    file_path = "/content/sample_data/california_housing_test.csv"

    loader = CSVLoader(file_path=file_path, encoding="utf-8", csv_args={'delimiter': ','})

    documents = loader.load()



    # Split documents into chunks

    text_splitter = CharacterTextSplitter(chunk_size=200, chunk_overlap=0)

    docs = text_splitter.split_documents(documents)



    print(docs)



    # Create the open-source embedding function

    embedding_function = SentenceTransformerEmbeddings(model_name="all-MiniLM-L6-v2")



    #persist_directory="./chroma_db"

    # Create Chroma index

    db2 = Chroma.from_documents(docs, embedding_function)



    # Configure LLM model

    llm_model_name = "TheBloke/Llama-2-7B-GGUF"




    llm = CTransformers(model=llm_model_name,

                        model_type='llama',

                        max_new_tokens=8145,

                        temperature=1.0)



    memory = ConversationBufferMemory(memory_key="chat_history", return_messages=False)



    # Create conversational retrieval chain

    qa = ConversationalRetrievalChain.from_llm(llm,

                                            retriever=db2.as_retriever(search_kwargs={'k': 10}),

                                            memory=memory,

                                            output_key="answer",

                                            verbose=True,

                                            max_tokens_limit=4000)



    # Interactive conversation loop

    while True:

        print(" ")

        query = input('Pergunta: ').lower()

        if query == 'quit':

            break

        print(" ")

        traducao = translate_to_portuguese(query)

        chat_history = []

        result = qa({"question": traducao, "chat_history": chat_history})

        response = GoogleTranslator(source="en", target="pt")

        traducao_response = response.translate(result['answer'])

        print("Resposta Bot:", traducao_response)



if __name__ == "__main__":

    main()  `

这个错误：

WARNING:ctransformers:Number of tokens (1554) exceeded maximum context length (512).

Answer 1

使用特定设置设置变压器模型并手动处理令牌。

例如：

llm = CTransformers(model=llm_model_name,
 model_type='llama',
 config={'max_new_tokens': 600, 'temperature': 0.01, 'context_length': 700})