Name: RAG vs 微調 vs 快速工程：優化人工智能模型 (RAG vs Fine-Tuning vs Prompt Engineering: Optimizing AI Models)
Uploaded: 2025-08-16T08:19:05.000Z
Duration: 13 min 10 s

反身代名詞

Remember how, back in the day, people would Google themselves, you'd type your name into a search engine, and you'd see what it knows about you?

還記得在過去，人們會用谷歌搜索自己嗎？在搜索引擎上輸入你的名字，就能看到它對你的瞭解。

Well, the modern equivalent of that is to do the same thing with a chatbot.

那麼，現代的等價物就是用哈拉機器人做同樣的事情。

程度副詞

So when I ask a large language model, who is Martin Kean, well, the response varies greatly depending upon which model I'm asking, because different models, they have different training datasets, they have different knowledge cutoff dates.

是以，當我問一個大型語言模型 "馬丁-基恩是誰 "時，不同模型的回答大相徑庭，因為不同的模型有不同的訓練數據集，有不同的知識截止日期。

So what a given model knows about me, well, it differs greatly.

是以，模特對我的瞭解程度也大相徑庭。

But how could we improve the model's answer?

但是，我們怎樣才能改進模型的答案呢？

So let's start with a model here, and we're going to see how we can improve its responses.

是以，讓我們從這裡的一個模型開始，看看如何改進它的響應。

Well, the first thing it could do is it could go out and it could perform a search, a search for new data that either wasn't in its training dataset, or it was just data that became available after the model finished training.

那麼，它可以做的第一件事就是出去執行搜索，搜索訓練數據集中沒有的新數據，或者是模型完成訓練後可用的數據。

And then it could incorporate those results from the search back into its answer.

然後，它就能把這些搜索結果反饋到它的答案中。

That is called RAG, or Retrieval Augmented Generation.

這就是所謂的 RAG，即 "檢索增強生成"。

Or we could pick a specialized model, a model that's been trained on, let's say, transcripts of these videos.

或者，我們可以選擇一個專門的模型，一個根據這些視頻的轉錄本訓練出來的模型。

That would be an example of something called fine tuning.

Or we could ask the model a query that better specifies what we're looking for.

或者，我們可以向模型提出一個查詢，讓它更好地說明我們要找的東西。

So maybe the LLM already knows plenty about the Martin Keens of the world, but let's tell the model that we're referring to the Martin Keen who works at IBM, rather than the Martin Keen that founded Keen Shoes.

是以，也許法律碩士已經對世界上的馬丁-基恩（Martin Keens）瞭如指掌，但讓我們告訴模型，我們指的是在 IBM 工作的馬丁-基恩（Martin Keen），而不是創辦基恩鞋業的馬丁-基恩（Martin Keen）。

That is an example of prompt engineering.

Three ways to get better outputs out of large language models, each with their pluses and minuses.

從大型語言模型中獲得更好輸出結果的三種方法，各有利弊。

So retrieval of external up-to-date information.

That's augmentation of the original prompt with the retrieved information added in.

這是在原有提示的基礎上增加了檢索到的資訊。

That's generation of a response based on all of this enriched context.

這就是在所有這些豐富的背景基礎上產生的反應。

So we start with a query, and the query comes in to a large language model.

是以，我們從一個查詢開始，然後將查詢輸入一個大型語言模型。

Now what RAG is going to do is it's first going to go searching through a corpus of information.

現在，RAG 要做的是首先搜索資訊語料庫。

So we have this corpus here full of some sort of data.

是以，我們這裡的語料庫充滿了某種數據。

Now perhaps that's your organisation's documents.

So it might be spreadsheets, PDFs, internal wikis, you know, stuff like that.

是以，它可能是電子表格、PDF、內部維基，你知道，諸如此類的東西。

But unlike a typical search engine that just matches keywords, RAG converts both your question, the query, and all of the documents into something called vector embeddings.

但與只匹配關鍵字的典型搜索引擎不同，RAG 會將問題、查詢和所有文檔轉換成一種叫做向量嵌入的東西。

So these are all converted into vectors, essentially turning words and phrases into long lists of numbers that capture their meaning.

是以，這些都被轉換成矢量，實質上就是將單詞和短語變成一長串數字，以捕捉其含義。

So when you ask a query like, what was our company's revenue growth last quarter?

是以，當你詢問 "我們公司上一季度的收入增長情況如何？

Well RAG will find documents that are mathematically similar in meaning to your question, even if they don't use the exact same words.

RAG 可以找到與您的問題在數學意義上相似的文檔，即使它們沒有使用完全相同的詞語。

So it might find documents mentioning fourth quarter performance or quarterly sales.

是以，它可能會找到提及第四季度業績或季度銷售額的文件。

Those don't contain the keyword revenue growth, but they are semantically similar.

這些內容不包含關鍵字收入增長，但它們在語義上是相似的。

Now once RAG finds the relevant information, it adds this information back into your original query before passing it to the language model.

現在，一旦 RAG 找到相關資訊，它就會將這些資訊添加回原始查詢中，然後再將其傳遞給語言模型。

So instead of the model just kind of guessing based on its training data, it can now generate a response that incorporates your actual facts and figures.

是以，該模型不再只是根據訓練數據進行猜測，而是能夠結合實際情況和數據生成響應。

So this makes RAG particularly valuable when you are looking for information that is up to date, and it's also very valuable when you need to add in information that is domain specific as well.

是以，當您要查找最新資訊時，RAG 就顯得尤為重要，而當您需要添加特定領域的資訊時，它也非常有價值。

So one cost, that would be the cost of performance for performing all of this, because you have this retrieval step here, and that adds latency to each query compared to a simple prompt to a model.

是以，其中一個成本就是執行所有這些操作的性能成本，因為這裡有檢索步驟，與簡單的模型提示相比，這會增加每次查詢的延遲時間。

There are also costs related to just kind of the processing of this as well.

此外，還需要支付相關的處理費用。

So if we think about what we are having to do here, we've got documents that need to be vector embeddings, and we need to store these vector embeddings in a database.

是以，如果我們考慮一下我們在這裡要做的事情，我們已經有了需要矢量嵌入的文檔，我們需要將這些矢量嵌入存儲到數據庫中。

All of this adds to processing costs, it adds to infrastructure costs to make this solution work.

所有這一切都增加了處理成本，增加了基礎設施成本，使這一解決方案無法發揮作用。

So remember how we discussed getting better answers about me by training a model specifically on let's say my video transcripts?

還記得我們討論過如何通過訓練一個模型來獲得關於我的更好答案嗎？

So what we do with fine tuning is we take a model, but specifically an existing model.

是以，我們所做的微調就是利用一個模型，特別是現有的模型。

And that existing model has broad knowledge.

而這種現有模式具有廣泛的知識。

And then we're going to give it additional specialised training on a focused data set.

然後，我們將在重點數據集上對它進行額外的專業培訓。

So this is now specialised to what we want to develop particular expertise on.

是以，現在的專業化是針對我們想要發展的特殊專長。

Now during fine tuning, we're updating the model's internal parameters through additional training.

現在，在微調過程中，我們通過額外的訓練來更新模型的內部參數。

So the model starts out with some weights here, kind of like this.

是以，模型一開始會有一些權重，有點像這樣。

And those weights were optimised during its initial pre-training.

而這些權重是在初始預訓練時優化的。

And as we fine tune, we're making small adjustments here to the model's weight using this specialised data set.

當我們進行微調時，我們會利用這組專業數據對模型的權重進行小幅調整。

Now this process typically uses supervised learning, where we provide input-output pairs that demonstrate the kind of responses we want.

現在，這一過程通常使用監督學習，我們提供輸入輸出對，展示我們想要的響應類型。

So for example, if we're fine tuning for technical support, we might provide thousands of examples of customer queries, and those would be paired with correct technical responses.

是以，舉例來說，如果我們要對技術支持進行微調，我們可能會提供成千上萬個客戶詢問的實例，並將這些實例與正確的技術回覆進行配對。

The model adjusts its weights through backpropagation to minimise the difference between its predicted outputs and the targeted responses.

該模型通過反向傳播調整權重，以儘量減少預測輸出與目標響應之間的差異。

So we're not just teaching the model new facts here, we're actually modifying how it processes information.

是以，我們在這裡不只是向模型傳授新知識，實際上是在改變它處理資訊的方式。

The model is learning to recognise domain-specific patterns.

該模型正在學習識別特定領域的模式。

So fine tuning shows its strength when you particularly need a model that has very deep domain expertise.

是以，當你特別需要一個擁有非常深厚領域專業知識的模型時，微調就會顯示出它的優勢。

That's what we can really add in with fine tuning.

這才是我們真正可以進行微調的地方。

And also, it's much faster, specifically at inference time.

而且，它的速度更快，特別是在推理時。

So when we are putting the queries in, it's faster than RAG because it doesn't need to search through external data.

是以，當我們輸入查詢時，它比 RAG 更快，因為它不需要搜索外部數據。

And because the knowledge is kind of baked into the model's weights, you don't need to maintain a separate vector database.

由於這些知識已經融入到模型的權重中，是以無需單獨維護向量數據庫。

Well, there's certainly issues here with the training complexity of all of this.

當然，培訓的複雜性也是個問題。

You're going to need thousands of high-quality training examples.

你需要成千上萬個高質量的培訓案例。

There are also issues with computational cost.

The computational cost for training this model can be substantial and it's going to require a whole bunch of GPUs.

訓練這個模型的計算成本可能很高，需要一大堆 GPU。

And there's also challenges related to maintenance as well, because unlike RAG, where you can easily add new documents to your knowledge base at any point, updating a fine-tuned model requires another round of training.

此外，在維護方面也存在挑戰，因為與 RAG 不同，RAG 可以隨時輕鬆地向知識庫中添加新文檔，而更新微調模型則需要另一輪培訓。

And then perhaps most importantly of all, there is a risk of something called catastrophic forgetting.

也許最重要的是，還有一種叫做災難性遺忘的風險。

Now that's when the model loses some of its general capabilities while it's busy learning these specialised ones.

這時，模型在忙於學習這些專業能力的同時，也失去了一些通用能力。

So finally, let's explore prompt engineering.

最後，讓我們來探討一下及時工程。

Specifying Martin Keen who works at IBM versus Martin Keen who founded Keen Shoes, that's prompt engineering, but at its most basic.

把在 IBM 工作的馬丁-基恩與創辦基恩鞋業的馬丁-基恩相提並論，這就是最基本的及時工程學。

Prompt engineering goes far beyond simple clarification.

及時的工程設計遠遠超出了簡單的澄清。

So let's think about when we input a prompt, the model receives this prompt and it processes it through a series of layers.

是以，讓我們想一想，當我們輸入一個提示時，模型會接收到這個提示，並通過一系列的層進行處理。

And these layers are essentially attention mechanisms.

而這些層次本質上就是注意力機制。

And each one focuses on different aspects of your prompt text that came in.

每篇文章都側重於您收到的提示文本的不同方面。

And by including specific elements in your prompt, so examples or context or how you want the format to look, you're directing the model's attention to relevant patterns it learned during training.

通過在提示中加入特定元素，如例子、上下文或您希望的格式，您可以引導模型關注它在訓練過程中學到的相關模式。

So for example, telling a model to think about this step by step, that activates patterns it learned from training data where methodical reasoning led to accurate results.

是以，舉例來說，告訴模型一步一步地思考這個問題，就能激活它從訓練數據中學到的模式，在這些數據中，有條不紊的推理能得出準確的結果。

So a well engineered prompt can transform a model's output without any additional training or without data retrieval.

是以，一個精心設計的提示可以在不進行任何額外訓練或數據檢索的情況下改變模型的輸出。

比方說，我們說，這段代碼安全嗎？

An engineered prompt, it might read a bit more like this.

如果是工程提示，可能會更像這樣。

We've just better activated its existing capabilities.

我們只是更好地激活了它的現有能力。

Now I think the benefits to this are pretty obvious.

現在，我認為這樣做的好處是顯而易見的。

One is that we don't need to change any of our back end infrastructure here because there are no infrastructure changes at all in order to prompt better.

一是我們不需要改變任何後端基礎設施，因為根本不需要改變基礎設施就能更好地進行提示。

There's also the benefit that by doing this, you get to see immediate responses and immediate results to what you do.

這樣做還有一個好處，那就是你可以看到對你所做事情的即時反應和即時結果。

We don't have to add in new training data or any kind of data processing.

我們無需添加新的訓練數據或進行任何數據處理。

But of course, there are some limitations to this as well.

Prompt engineering is as much an art as it is a science.

快速工程既是一門藝術，也是一門科學。