很好的LLM知识博客:
https://lilianweng.github.io/posts/2023-06-23-agent/
LangChain的prompt hub:
https://smith.langchain.com/hub
一. Q&A
1. Q&A
os.environ["OPENAI_API_KEY"] = “OpenAI的KEY” # 把openai-key放到环境变量里;
from langchain_openai import ChatOpenAI
llm = ChatOpenAI(model="gpt-4-0125-preview")
text_splitter = RecursiveCharacterTextSplitter(chunk_size=1000, chunk_overlap=200) # 相邻chunk之间是200字符。(可以指定按照\n还是句号来分割)
splits = text_splitter.split_documents(docs) # 文档切块
vectorstore = Chroma.from_documents(documents=splits, embedding=OpenAIEmbeddings()) # 使用openai的embedding来对文档做向量化;使用Chroma向量库;
retriever = vectorstore.as_retriever()
prompt = hub.pull("rlm/rag-prompt") # 从hub上拉取现成的prompt
rag_chain = (
{"context": retriever | format_docs, "question": RunnablePassthrough()}
| prompt
| llm
| StrOutputParser()
) # 用竖线"|"来串联起各个组件;
2. Chat history
用LangChain封装好的组件,把2个子chain连接在一起。
Query改写的目的:
Q1:"任务分解指的是什么?"
A: "..."
Q2: "它分为哪几步?"
有了Query改写,Q2可被改写为"任务分解分为哪几步?", 进而使用其作为query查询到有用的doc;
2.5 Serving
和FastAPI集成:
from fastapi import FastAPI
from langserve import add_routes# 4. Create chain
chain = prompt_template | model | parser# 4. App definition
app = FastAPI(title="LangChain Server",version="1.0",description="A simple API server using LangChain's Runnable interfaces",
)# 5. Adding chain routeadd_routes(app,chain,path="/chain",
)if __name__ == "__main__":import uvicornuvicorn.run(app, host="localhost", port=8000)
3. Streaming
LangChain支持流式输出。
4.1 Wiki
langchain提供现成的Wikipedia数据库;已经向量化完毕封装在类里了。
from langchain_community.retrievers import WikipediaRetriever
wiki = WikipediaRetriever(top_k_results=6, doc_content_chars_max=2000)
4. Citation
让模型给出answer来自于context中docs的哪些片段。
可以在prompt里让模型给出answer的同时,也给出引用doc的id和文字片段。
例如:(注意"VERBATIM", 一字不差的)
You're a helpful AI assistant. Given a user question and some Wikipedia article snippets, \
answer the user question and provide citations. If none of the articles answer the question, just say you don't know.Remember, you must return both an answer and citations. A citation consists of a VERBATIM quote that \
justifies the answer and the ID of the quote article. Return a citation for every quote across all articles \
that justify the answer. Use the following format for your final output:<cited_answer><answer></answer><citations><citation><source_id></source_id><quote></quote></citation><citation><source_id></source_id><quote></quote></citation>...</citations>
</cited_answer>Here are the Wikipedia articles:{context}"""
prompt_3 = ChatPromptTemplate.from_messages([("system", system), ("human", "{question}")]
)
所谓的tool,原理可能(我猜)也是tool封装了以上的改prompt方法。
也可以先调LLM给出answer,再调一次LLM给出citation。缺点是调用了2次LLM。
二. structured output
1. LangChain支持一种叫做Pydantic的语法,描述output:
from typing import Optional
from langchain_core.pydantic_v1 import BaseModel, Fieldclass Person(BaseModel):"""Information about a person."""# ^ Doc-string for the entity Person.# This doc-string is sent to the LLM as the description of the schema Person,# and it can help to improve extraction results.# Note that:# 1. Each field is an `optional` -- this allows the model to decline to extract it!# 2. Each field has a `description` -- this description is used by the LLM.# Having a good description can help improve extraction results.name: Optional[str] = Field(default=None, description="The name of the person")hair_color: Optional[str] = Field(default=None, description="The color of the peron's hair if known")height_in_meters: Optional[str] = Field(default=None, description="Height measured in meters")
其中,加了"Optional"的,有则输出,无则不输出。
runnable = prompt | llm.with_structured_output(schema=Person)
text = "Alan Smith is 6 feet tall and has blond hair."
runnable.invoke({"text": text})
注意:该llm必须是支持该Pydantic和with_structured_output的模型才行。
输出结果:
Person(name='Alan Smith', hair_color='blond', height_in_meters='1.8288')
2. 提升效果的经验
- Set the model temperature to
0
. (只拿最优解) - Improve the prompt. The prompt should be precise and to the point.
- Document the schema: Make sure the schema is documented to provide more information to the LLM.
- Provide reference examples! Diverse examples can help, including examples where nothing should be extracted. (Few shot例子!)
- If you have a lot of examples, use a retriever to retrieve the most relevant examples. (Few shot例子多些更好;正例和负例都要覆盖到,负例是指抽取不到所需字段的情况)
- Benchmark with the best available LLM/Chat Model (e.g., gpt-4, claude-3, etc) -- check with the model provider which one is the latest and greatest! (有的模型连结构化格式都经常输出得不对。。。最好专门用结构化输出来做训练,再用)
- If the schema is very large, try breaking it into multiple smaller schemas, run separate extractions and merge the results. (一次输出的格式,不能太复杂;否则要拆分)
- Make sure that the schema allows the model to REJECT extracting information. If it doesn't, the model will be forced to make up information! (提示模型,可以拒绝输出,不懂不要乱说)
- Add verification/correction steps (ask an LLM to correct or verify the results of the extraction). (用大模型或者代码或者人工来验证正确性,json load失败或不包含必要字段,则说明输出格式错误)
三. Chatbot
1. 有base model和chat model,一定要选用专门为chat做过训练的chat model!
2. 两处query改写:
带知识库检索的,要注意query改写,将类似"那是什么”中的"那"这种代词给替换掉,再去检索。
带memory的,也要query改写。
3. 精简memory的目的:A. 大模型context长度有限;B.删去对话历史中的无关内容,让大模型更聚焦在有用信息上。
memory总结用的prompt:
Distill the above chat messages into a single summary message. Include as many specific details as you can
4. 知识检索,要记得加上“不懂别乱说”:
If the context doesn't contain any relevant information to the question, don't make something up and just say "I don't know"
四. Tools&Agent
1. Agent:大模型来决定,这一步用什么tool(or 直接输出最终结果)以及入参;每一步都把上一步的输出和tool的输出作为历史信息。
2. LangChain支持自定义tool:
from langchain_core.tools import tool@tool
def multiply(first_int: int, second_int: int) -> int:"""Multiply two integers together."""return first_int * second_int
注释很有用,告诉大模型该tool是干什么用的。
3. tool调用失败后的策略:
A. 给一个兜底LLM(GPT4等能力更强的模型),第一个模型失败后,调兜底模型再试。
B. 把tool入参和报错信息,放到prompt里,再调用一次(大模型很听话)。
"The last tool call raised an exception. Try calling the tool again with corrected arguments. Do not repeat mistakes."
五. query & analysis
1. 如果知识库是结构化/半结构化数据,可以把query输入大模型得到更合适的搜索query,例如{"query":"XXX", "publish_year":"2024"}
2. 一个复杂query,输入LLM,得到若干简单query;使用简单query查询知识库得到结果,合在一起,回答复杂query
帮助LLM理解如何去分解得到简单query? 答:给几个few-shot-examples;
3. "需要搜索则输出query,不需要搜索则直接输出回答"
4. 多个知识库:根据生成的query里的“库”字段,只查询相应的库;没必要所有库都查询;