Making large AI models cheaper, faster and more accessible
You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
 
 
 
 
 

113 lines
4.4 KiB

"""
Script for Chinese retrieval based conversation system backed by ChatGLM
"""
import argparse
import os
from colossalqa.chain.retrieval_qa.base import RetrievalQA
from colossalqa.data_loader.document_loader import DocumentLoader
from colossalqa.local.llm import ColossalAPI, ColossalLLM
from colossalqa.memory import ConversationBufferWithSummary
from colossalqa.prompt.prompt import (
PROMPT_DISAMBIGUATE_ZH,
PROMPT_RETRIEVAL_QA_ZH,
SUMMARY_PROMPT_ZH,
ZH_RETRIEVAL_QA_REJECTION_ANSWER,
ZH_RETRIEVAL_QA_TRIGGER_KEYWORDS,
)
from colossalqa.retriever import CustomRetriever
from colossalqa.text_splitter import ChineseTextSplitter
from langchain import LLMChain
from langchain.embeddings import HuggingFaceEmbeddings
if __name__ == "__main__":
# Parse arguments
parser = argparse.ArgumentParser(description="Chinese retrieval based conversation system backed by ChatGLM2")
parser.add_argument("--model_path", type=str, default=None, help="path to the model")
parser.add_argument("--model_name", type=str, default=None, help="name of the model")
parser.add_argument(
"--sql_file_path", type=str, default=None, help="path to the a empty folder for storing sql files for indexing"
)
args = parser.parse_args()
if not os.path.exists(args.sql_file_path):
os.makedirs(args.sql_file_path)
colossal_api = ColossalAPI.get_api(args.model_name, args.model_path)
llm = ColossalLLM(n=1, api=colossal_api)
# Setup embedding model locally
embedding = HuggingFaceEmbeddings(
model_name="moka-ai/m3e-base", model_kwargs={"device": "cpu"}, encode_kwargs={"normalize_embeddings": False}
)
# Define the retriever
information_retriever = CustomRetriever(k=3, sql_file_path=args.sql_file_path, verbose=True)
# Define memory with summarization ability
memory = ConversationBufferWithSummary(
llm=llm,
prompt=SUMMARY_PROMPT_ZH,
human_prefix="用户",
ai_prefix="Assistant",
max_tokens=2000,
llm_kwargs={"max_new_tokens": 50, "temperature": 0.6, "do_sample": True},
)
# Define the chain to preprocess the input
# Disambiguate the input. e.g. "What is the capital of that country?" -> "What is the capital of France?"
llm_chain_disambiguate = LLMChain(
llm=llm, prompt=PROMPT_DISAMBIGUATE_ZH, llm_kwargs={"max_new_tokens": 30, "temperature": 0.6, "do_sample": True}
)
def disambiguity(input: str):
out = llm_chain_disambiguate.run(input=input, chat_history=memory.buffer, stop=["\n"])
return out.split("\n")[0]
# Load data to vector store
print("Select files for constructing retriever")
documents = []
while True:
file = input("Enter a file path or press Enter directory without input to exit:").strip()
if file == "":
break
data_name = input("Enter a short description of the data:")
retriever_data = DocumentLoader([[file, data_name.replace(" ", "_")]]).all_data
# Split
text_splitter = ChineseTextSplitter()
splits = text_splitter.split_documents(retriever_data)
documents.extend(splits)
# Create retriever
information_retriever.add_documents(docs=documents, cleanup="incremental", mode="by_source", embedding=embedding)
# Set document retrieval chain, we need this chain to calculate prompt length
memory.initiate_document_retrieval_chain(llm, PROMPT_RETRIEVAL_QA_ZH, information_retriever)
# Define retrieval chain
llm_chain = RetrievalQA.from_chain_type(
llm=llm,
verbose=False,
chain_type="stuff",
retriever=information_retriever,
chain_type_kwargs={"prompt": PROMPT_RETRIEVAL_QA_ZH, "memory": memory},
llm_kwargs={"max_new_tokens": 150, "temperature": 0.6, "do_sample": True},
)
# Set disambiguity handler
information_retriever.set_rephrase_handler(disambiguity)
# Start conversation
while True:
user_input = input("User: ")
if "END" == user_input:
print("Agent: Happy to chat with you :)")
break
agent_response = llm_chain.run(
query=user_input,
stop=["</答案>"],
doc_prefix="支持文档",
rejection_trigger_keywords=ZH_RETRIEVAL_QA_TRIGGER_KEYWORDS,
rejection_answer=ZH_RETRIEVAL_QA_REJECTION_ANSWER,
)
print(f"Agent: {agent_response}")