超越知识截止
大型语言模型功能强大,但存在一个根本性局限: 知识截止。为了构建可靠的AI系统,我们必须弥合静态训练数据与动态现实信息之间的差距。
1. 知识截止问题(是什么)
大语言模型在大规模但静态的数据集上进行训练,这些数据有固定的截止日期(例如,GPT-4的截止时间为2021年9月)。因此,模型无法回答关于近期事件、软件更新或训练完成后创建的私有数据的问题。
2. 幻觉与现实(为什么)
当被问及未知或截止日期后的数据时,模型常常 产生幻觉——编造听起来合理但实际上完全错误的事实以满足提示要求。解决方案是 信息锚定:在模型生成答案前,从外部知识库提供实时且可验证的上下文。
3. RAG 与 微调(如何)
- 微调: 更新模型内部权重计算成本高、速度慢,并导致知识静态化,很快就会过时。
- RAG (检索增强生成): 成本极低。它能即时检索相关信息并注入到提示中,确保数据最新,并可在不重新训练的情况下轻松更新知识库。
私有数据鸿沟
除非通过检索管道显式集成,否则大语言模型无法访问公司内部手册、财务报告或机密文件。
TERMINALbash — 80x24
> Ready. Click "Run" to execute.
>
Question 1
Why is Retrieval Augmented Generation (RAG) preferred over fine-tuning for updating an LLM's knowledge of daily news?
Question 2
What term describes an LLM's tendency to invent facts when it lacks information?
Challenge: Building a Support Bot
Apply RAG concepts to a real-world scenario.
You are building a support bot for a new product released today. The LLM you are using was trained two years ago.
Task 1
Identify the first step in the RAG pipeline to get the product manual into the system so the LLM can search it.
Solution:
Preprocessing (Cleaning and chunking the manual text into smaller, searchable segments before embedding).
Preprocessing (Cleaning and chunking the manual text into smaller, searchable segments before embedding).
Task 2
Define a "System Message" that forces the LLM to only use the provided documents and prevents hallucination.
Solution:
"Answer only using the provided context. If the answer is not in the context, state that you do not know."
"Answer only using the provided context. If the answer is not in the context, state that you do not know."