以上为实现效果
RAG是目前最火的大模型应用之一，如何能快速实现一个不错的demo呢？
参考
- https://github.com/LongxingTan/open-retrievals
- https://colab.research.google.com/drive/1fJC-8er-a4NRkdJkwWr4On7lGt9rAO4P?usp=sharing#scrollTo=2Hrfp96UYufn
出于硬件考虑，目前用的大模型是Qwen1.5最小的size，考虑到涌现至少需要10B以上的大模型，有更好的硬件还可以表现的更好
如何确定大模型是否使用资料呢？根据应用类型有几种思路。是否是一直从一个常备知识库中进行查询的应用？还是每次都上传一个文件，根据文件进行问答后，多轮问答后即舍弃该知识. 或者由大模型判断问题是否需要进一步的知识再从知识库中召回
代码
主要来自
https://github.com/LongxingTan/open-retrievals/blob/master/examples/3_rag/rag_langchain_demo.py
import json
import os
import re
import tempfile
import streamlit as st
from langchain import LLMChain, PromptTemplate
from langchain.chains import RetrievalQA
from langchain.document_loaders import PyPDFLoader
from langchain.memory.chat_message_histories import StreamlitChatMessageHistory
from langchain.retrievers import EnsembleRetriever
from langchain.text_splitter import RecursiveCharacterTextSplitter
from langchain.vectorstores import Chroma
from retrievals.tools.langchain import LangchainEmbedding, LangchainLLM
st.set_page_config(page_title="RAG with Open-retrievals")
with st.sidebar:
st.write("**RAG with Open-retrievals**")
@st.cache_resource(ttl="1h")
def configure_retriever(uploaded_files):
# read documents
docs = []
temp_dir = tempfile.TemporaryDirectory()
for file in uploaded_files:
temp_filepath = os.path.join(temp_dir.name, file.name)
with open(temp_filepath, "wb") as f:
f.write(file.getvalue())
loader = PyPDFLoader(temp_filepath)
docs.extend(loader.load())
# split documents
text_splitter = RecursiveCharacterTextSplitter(chunk_size=500, chunk_overlap=50)
splits = text_splitter.split_documents(docs)
embeddings = LangchainEmbedding(model_name_or_path="sentence-transformers/all-MiniLM-L6-v2")
vectordb = Chroma.from_documents(splits, embeddings)
retrieval_args = {"search_type": "similarity", "score_threshold": 0.15, "k": 30}
retriever = vectordb.as_retriever(**retrieval_args)
return retriever
uploaded_files = st.sidebar.file_uploader(label="Upload PDF files", type=["pdf"], accept_multiple_files=True)
if not uploaded_files:
st.info("Please upload PDF documents to continue.")
st.stop()
retriever = configure_retriever(uploaded_files)
llm = LangchainLLM(model_name_or_path="Qwen/Qwen1.5-1.8B-Chat", temperature=0.5, max_tokens=2048, top_k=10)
msgs = StreamlitChatMessageHistory()
RESPONSE_TEMPLATE = """[INST]
<>
You are a helpful AI assistant. Use the following pieces of context to answer the user's question.<>
Anything between the following `context` html blocks is retrieved from a knowledge base.
{context}
REMEMBER:
- If you don't know the answer, just say that you don't know, don't try to make up an answer.
- Let's take a deep breath and think step-by-step.
Question: {question}[/INST]
Helpful Answer:
"""
PROMPT = PromptTemplate.from_template(RESPONSE_TEMPLATE)
PROMPT = PromptTemplate(template=RESPONSE_TEMPLATE, input_variables=["context", "question"])
qa_chain = RetrievalQA.from_chain_type(
llm,
chain_type='stuff',
retriever=retriever,
chain_type_kwargs={
"verbose": True,
"prompt": PROMPT,
},
)
if len(msgs.messages) == 0 or st.sidebar.button("New Chat"):
msgs.clear()
msgs.add_ai_message("How can I help you?")
avatars = {"human": "user", "ai": "assistant"}
for msg in msgs.messages:
st.chat_message(avatars[msg.type]).write(msg.content)
if user_query := st.chat_input(placeholder="Ask me anything!"):
st.chat_message("user").write(user_query)
with st.chat_message("assistant"):
response = qa_chain({"query": user_query})
answer = response["result"]
st.write(answer)
about = st.sidebar.expander("About")
about.write("Powered by [open-retrievals](https://github.com/LongxingTan/open-retrievals)")
Dense 解读
- https://github.com/luyug/Dense
example中的wiki数据
- Dataset:
单个样本, (query, [positive, negative])
query是：{‘input_ids’: query_tensor}
passage是：[{‘input_ids’: pos_tensor}, {‘input_ids’: neg_tensor}]
经过collator后: pad
- Model
encode query:
encode passage:
损失函数:
