剖析 GPT 第三篇: 隱式知識 vs 顯式記憶 Dissecting GPT, Part 3: Implicit Knowledge vs. Explicit Memory I

隱式知識 vs 顯式記憶:

為什麼真正好用的 AI 一定長得像「顧問+資料室」組合?

前言:當 AI 顧問講得頭頭是道,卻講錯數字



先從一個你大概已經遇過的場景開始。

你把公司的一份簡報丟給 ChatGPT,問它:

「幫我整理這家公司過去三年的成長關鍵,順便補一些產業背景。」

AI 給你一份整理得非常漂亮的摘要:

  • 有邏輯、有結構

  • 用詞專業、語氣沉穩

  • 看起來完全像是某位顧問公司副理寫出來的東西

直到你開始對照實際數字,才發現怪怪的:

  • 營收成長率講錯

  • EBITDA margin 跟實際財報不一樣

  • 引用的市場規模數字也不知道是哪一年的

內容「有 sense」,但細節「不靠譜」。

如果你有這樣的體驗,其實不是你運氣不好,而是 LLM 的設計本來就如此:

它天生擅長的是「理解與表達」,
不是「記住最新、最精準的事實」。

於是問題來了:

  • 要怎麼讓這個超強「語言顧問」,擁有一個永遠更新中的資料室?

  • 要怎麼讓它在講故事之前,先真的讀過相關文件?

這就是我們今天要談的主題:
隱式知識 vs 顯式記憶,還有 RAG、GraphRAG 這些名字背後的真實角色。


一、LLM 的隱式知識:一個見多識廣、卻不會背財報的顧問

在前兩篇裡,我們已經把 LLM 講成一個東西:

它不是資料庫,而是一個用大量文本訓練出來的「世界模型」。
知識被壓在權重裡,以分佈式的方式存在,變成一種「數位直覺」。

這種直覺有幾個很明顯的優點:

  • 看得懂文本:你丟給它一篇報告,它能快速抓到重點

  • 會整合與重寫:可以幫你把散落的資訊重組成一個有邏輯的故事

  • 有推理能力:能回答「為什麼」、「如果…會怎樣」這種問題

  • 語氣可調整:你叫它寫成簡報語氣、新聞稿語氣、投資備忘錄語氣,它都能配合

聽起來幾乎完美,但真正用在工作就會發現它有幾個天然弱點:

  1. 不擅長記「最新」的事
    模型的知識停留在它最後一次訓練的時間點之後。
    過去半年發生了什麼,它未必知道。

  2. 對精準數字與條文不牢靠
    要它講故事可以,要它背 2023 Q4 的毛利率就很危險。

  3. 對公司內部知識完全一無所知
    SOP、內部制度、專案文件、內部數據報表,通通不在它的訓練資料裡。

所以,把 LLM 單獨丟進企業裡期待它「什麼都能回答」,
有點像請一位麥肯錫顧問來公司,
卻不給他任何內部資料,叫他憑印象說說看。

他講出來的策略方向可能很有道理,
但你一定不會放心讓他「亂填財報細節」。


二、顯式記憶:向量資料庫、全文檢索、GraphRAG 到底差在哪裡?

為了補上 LLM 的天然短板,工程師們又搬出了另一整個世界:顯式記憶系統

這一類技術聽起來名詞很多,但關鍵角色其實不難理解。

1. 全文檢索:關鍵字時代的資深圖書管理員

這是最傳統的檢索方式。

  • 你輸入關鍵字:如「利率決策」、「台積電 2023 法說」

  • 系統在一堆文件裡找出「文字上」匹配的段落

  • 再依照關聯度排序給你看

這很像一位老派圖書館員擅長的事:
你說出主題,他知道哪幾櫃有相關書。

優點:

  • 靠關鍵字精準匹配

  • 實作成熟、工具多

缺點也很明顯:

  • 你得自己會「下關鍵字」

  • 不同說法可能找不到一樣的內容(語意不同但本質相同的問題)

2. 向量資料庫:懂一點語意的新時代資料室

向量資料庫(Vector DB)做了一點升級:

  • 不再只比對「字長得像不像」

  • 而是把每段文字轉成「語意向量」

  • 再用向量距離衡量「意思像不像」

例如:

  • 「日圓長期貶值對出口有什麼影響?」

  • 跟「匯率偏低是否有利於日本出口?」

字眼不同,但意思相近。
向量資料庫可以把這兩者拉得比關鍵字搜尋更接近。

你可以把它想成:

一位比較懂內容的圖書管理員,
不只看書名,還大概知道書裡在講什麼。

3. GraphRAG:從「一堆段落」變成「節點與關係」

再往上一層,是 GraphRAG 這一類技術。
它不只關心「哪一段文字相關」,還關心:

  • 這些段落裡出現的「實體」(公司、人、產品、事件)

  • 實體與實體之間的「關係」(併購、合作、競爭、上下游)

這就把資料從「一堆段落」,變成「一張圖」。

用商業語言來說:

GraphRAG 比較像是「策略分析師」會畫的「產業關係圖」、「價值鏈地圖」。

你不只是知道:

  • A 公司提到 B 產品幾次

而是能看到:

  • A 公司是 B 產品的代工廠

  • B 產品主要銷售給 C 品牌

  • C 品牌近期又跟 D 產線互動頻繁

這類圖狀結構,對很多問題會比「一堆文字」更有用。


三、為什麼光有 LLM 或光有 RAG,都不夠?

如果你只用 LLM:

  • 會講故事

  • 會推理

  • 會寫得很順

  • 但容易講錯具體事實、年份、數字

如果你只用 RAG(或只有搜尋系統):

  • 能找到正確的段落

  • 能保證「原文」不會寫錯

  • 但不會幫你解讀:什麼重要?彼此有什麼關係?代表什麼趨勢?

用一句話總結:

LLM 擅長「理解與生成」,不擅長「記錄與更新」。
RAG 擅長「記錄與查找」,完全不會「理解與推理」。

真正強的,是兩者合體之後。


四、RAG 的本質:讓 AI 顧問在說話前,先把資料讀過一遍

RAG 全名是 Retrieval-Augmented Generation,
直譯叫「檢索增強生成」。

工程名詞聽起來有點抽象,把它翻成「日常工作流程」就很清楚:

想像你是一位策略顧問,要寫一份產業簡報

正常流程不會是:

直接坐下來,憑印象開始打字。

而是:

  1. 先到資料室(或內部知識庫)

    • 把過去的研究報告翻出來

    • 把最近幾季的財報掃過

    • 看一看產業新聞與重點事件

  2. 整理出幾份「最相關」的資料放桌上

  3. 再開始:

    • 抽取重點

    • 做比較

    • 下結論

    • 寫成一份給決策者看的簡報

RAG 做的事情,其實就對應到這三步:

  1. Retrieval(檢索)
    用向量資料庫、全文檢索、GraphRAG 等工具,
    幫你從大量文件裡挑出「跟問題最相關的部分」。

  2. Augmentation(增強)
    把這些段落、關係圖、數據,
    當作「Context(上下文)」塞給 LLM。

  3. Generation(生成)
    讓 LLM 在「看過這些資料之後」,
    用自己的隱式世界模型來:

    • 解讀

    • 重寫

    • 推理

    • 做總結

你可以把差別理解為:

沒有 RAG:
「我大概知道這個產業的共通套路,我講一個合理的版本給你聽。」

有 RAG:
「我先看過你指定的資料,再用我對世界的理解幫你整理出一份專屬的分析。」


五、GraphRAG:當問題不再只是「找幾段文字」,而是「找出關係網」

普通的 RAG,大多數時候是「找幾段相關的文件」。
但很多商業問題,根本不是單一文件可以回答的。

舉幾個例子:

  • 「這三家公司在供應鏈上的位置怎麼互動?」

  • 「這幾年來,哪幾件併購案改變了產業格局?」

  • 「在我們內部專案歷史裡,哪幾個決策節點後來證明是關鍵?」

這些問題,比較像是要:

  • 找出節點(公司、專案、產品、事件)

  • 看清節點之間的連結方式

  • 再由 LLM 幫你說出「這張圖代表什麼」

GraphRAG 做的,就是在 RAG 之前多一個步驟:

  1. 先從大量文件裡抽取出:

    • 人名、公司名、專案名、產品名、時間、地點

    • 以及它們間的關係(投資、合作、競爭、上下游…)

  2. 把這些變成一張「圖」(graph)

    • 節點是實體

    • 邊是關係

  3. 當你提問時,

    • 不是只找「相似段落」,

    • 而是沿著圖上的關係路徑,找到跟你問題最相關的一小塊子圖。

  4. 最後,再把這些子圖轉成 LLM 看得懂的文字描述(或結構化資料),
    交給它做解讀。

等於是:

普通 RAG:從一大堆文章裡,挑出幾個應該要看的段落。
GraphRAG:從整張關係圖裡,挑出幾條你現在最該關心的線。


六、對企業來說,這套「顧問+資料室」組合可以落地在哪?

用這種思路看,你會發現很多應用都可以重新改寫。

1. 內部知識庫問答

情境:
新員工問:「我們公司對某類客戶的報價策略是什麼?」

做法:

  • RAG 從內部文件、簡報、會議紀錄中找出相關內容

  • LLM 把它們整合成一份

    • 條列清楚

    • 有版本脈絡

    • 顯示注意事項的回答

好處:

  • 答案不是「憑印象編」,而是「有憑有據再加上解讀」。

2. 投資研究/產業掃描

情境:
你想快速了解一個新產業,或追蹤一段時間內的變化。

組合拳:

  • 外部:RAG / GraphRAG 搜新聞、研究報告、官方資料、法說會紀錄

  • 內部:RAG 搜自家過去寫過的投資備忘錄、模型、會議記錄

  • LLM:幫你統整成:

    • 產業結構

    • 主要玩家

    • 關鍵風險

    • 可能的情境分析

這比只看搜尋結果或只問 ChatGPT,都來得穩健。

3. 法遵、合約、政策解讀

情境:
法務或 compliance 團隊想知道「新法規對公司有哪些具體影響?」

流程:

  1. RAG 找出相關條文、本地解釋、公會意見、內部法遵文件

  2. LLM:

    • 幫你對照新舊規範

    • 整理「必須修改的內部流程」

    • 匯總成給管理層看的摘要

關鍵是:
條文本身仍來自原始文件,LLM 負責的是解釋與整理。


七、設計這種系統時,最容易踩的幾個坑

很多團隊在導入 RAG + LLM 時,會遇到一些常見問題。

1. 以為「丟給 LLM 自己去找就好」

這種設計的問題是:

  • 你永遠不知道它到底「有沒有真的看到該看的文件」

  • 你無法控制它用的是不是最新版本

  • 出錯時也很難回溯「錯在哪一段資料」

比較健康的做法是:

檢索(RAG)是顯式、可檢查的;
生成(LLM)是有依據、可追溯的。

2. 只把文件丟進向量資料庫,不做結構化整理

純文字當然可以用,但有些資料天生更適合結構化表示:

  • 時間線(timeline)

  • 關係圖(graph)

  • 表格數據(表頭+欄位)

GraphRAG 的價值往往就出現在這裡:
先把「圖」整理好,再讓 LLM 解讀圖,而不是直接淹沒在文字裡。

3. 沒有清楚劃分「誰負責事實,誰負責解釋」

一個簡單但重要的原則:

  • 「事實來源」要可追溯(某段文字、某個文件、某行資料表)

  • 「結論與建議」可以交給 LLM 用它的世界模型來生產

這樣錯了,你還找得到是哪邊的事實需要更新,
而不是整個系統看起來像在「亂說」。


小結:真正強大的 AI 系統,長得比較像「團隊」,而不是「單一模型」

如果要用一句話總結這一篇:

LLM 是一個見多識廣、會推理、會說話的顧問;
RAG、GraphRAG、向量資料庫是它的資料室與研究助理。

  • 隱式知識(模型權重):帶來理解、推理、風格、一致性

  • 顯式記憶(外部檢索):帶來即時性、精準度、可追溯性

兩者合體後,你才有一個:

  • 能讀文件、

  • 會查資料、

  • 懂得解釋、

  • 可以溝通、

  • 並且在犯錯時有跡可循的 AI 夥伴。

而這只是在講「模型怎麼用」。

下一篇,我們要往另一個方向走:
模型本身是怎麼「長大」的?

  • 預訓練像主力「吸籌」:先把世界規律買滿,建立底部

  • 微調像「洗盤」:調整風格與方向,去掉不必要的雜訊

  • RLHF 像「拉升」:讓能力以符合人類偏好的方式展現

我們會用完整一篇,把「Pretrain / Fine-tune / RLHF」
講成一套股市操作的故事。
========

本著作由大力士的AI天地創造 製作,以創用CC 姓名標示–非商業性– 禁止改作 4.0 國際授權條款釋出。


Dissecting GPT, Part 3: Implicit Knowledge vs. Explicit Memory

Implicit Knowledge vs. Explicit Memory:
Why truly useful AI looks like a “consultant + records room” combo

Preface: When the AI consultant sounds brilliant—but gets the numbers wrong

Let’s start with a situation you’ve probably encountered.

You drop a company slide deck into ChatGPT and ask:

“Help me summarize this company’s growth drivers over the past three years, and add some industry context.”

The AI returns a polished summary:

- Logical and well-structured
- Professional wording, steady tone
- Reads exactly like something a consulting firm manager would write

Until you check the actual figures and notice issues:

- Revenue growth rate is wrong
- EBITDA margin doesn’t match the financials
- The market size it cites isn’t from the year you’d expect

The content “has good sense,” but the details “aren’t reliable.”

If you’ve had this experience, it’s not bad luck—it’s how LLMs are built:

They’re naturally strong at “understanding and expression,”
not at “recalling the newest, most precise facts.”

Which raises two questions:

- How do we give this super “language consultant” a constantly updated records room?
- How do we make it actually read the relevant documents before telling the story?

That’s today’s topic:
Implicit knowledge vs. explicit memory—and the real roles behind RAG and GraphRAG.

1) The LLM’s implicit knowledge: a well-read consultant who won’t memorize your P&L

In Parts 1 and 2, we framed the LLM as:

Not a database, but a “world model” trained on massive text.
Knowledge is compressed into weights as distributed representations—a kind of “digital intuition.”

That intuition has clear strengths:

- Text comprehension: throw it a report and it quickly finds the core ideas
- Synthesis and rewriting: it recombines scattered information into a coherent narrative
- Reasoning ability: it can answer “why” and “what if” questions
- Tone control: it can write like a slide deck, press release, or investor memo on cue

Yet in real work, a few weaknesses show up:

- Poor at “the latest”
  The model’s knowledge freezes at its last training cutoff.
  It may not know what happened in the past six months.

- Unreliable for exact numbers and legal text
  Great at storytelling; risky for “what was 2023 Q4 gross margin?”

- Zero awareness of your internal knowledge
  SOPs, internal policies, project docs, internal dashboards—none are in the training data.

So expecting an LLM to “answer everything” in a company
is like hiring a McKinsey consultant
and giving them no internal materials—just “go off your memory.”

They’ll produce thoughtful strategic angles,
but you wouldn’t trust them to “fill in the financial footnotes.”

2) Explicit memory: what’s the difference between full-text search, vector databases, and GraphRAG?

To complement the LLM’s natural blind spots, engineers bring in a whole other universe: explicit memory systems.

The names sound many, but the roles are straightforward.

1. Full-text search: the veteran, keyword-era librarian

This is the classic approach.

- You enter keywords like “rate decision,” “TSMC 2023 earnings call”
- The system finds “textually” matching passages across documents
- Results are ranked by relevance

It’s like an old-school librarian:
you say the topic; they know which shelves to point to.

Pros:
- Fast
- Precise keyword matching
- Mature tools and implementations

Cons:
- You must craft good keywords
- Different phrasings can miss the same concept (semantic equivalents slip by)

2. Vector databases: a modern records room that understands some semantics

Vector DBs add a layer:

- They don’t just compare surface words
- Each passage is embedded into a “semantic vector”
- Similarity is measured by vector distance (semantic closeness)

For example:
“Long-term yen depreciation—what does it mean for exports?”
vs.
“Does a weak exchange rate benefit Japanese exporters?”

Different wording, similar meaning.
Vector search will pull them closer than pure keywords.

Think of it as:
a librarian who doesn’t just read titles—roughly understands what’s inside.

3. GraphRAG: from “a pile of passages” to “entities and relationships”

One level up is GraphRAG.
It cares not only about “which passages are relevant,” but also about:

- The “entities” inside those passages (companies, people, products, events)
- The “relationships” between entities (M&A, partnerships, competition, supply chain)

This turns data from “many paragraphs” into “a graph.”

In business terms:
GraphRAG is like the “industry maps” and “value chain diagrams” a strategy analyst draws.

You move from:
- “Company A mentions Product B X times”
to:
- “Company A is a contract manufacturer for Product B”
- “Product B primarily sells to Brand C”
- “Brand C has recently increased activity with Line D”

For many problems, graph structure is more useful than raw text.

3) Why LLM-only or RAG-only is not enough

LLM-only:
- Great storytelling
- Solid reasoning
- Smooth writing
- But likely to fumble specific facts, years, or numbers

RAG-only (or search-only):
- Finds the right passages
- Guarantees the “original text” is accurate
- But won’t interpret: what matters? what relates to what? what trend does it imply?

In one line:

LLMs excel at “understanding and generating,” not “recording and updating.”
RAG excels at “recording and retrieval,” not “understanding and reasoning.”

The real power emerges when the two are combined.

4) The essence of RAG: make the AI consultant read the materials before speaking

RAG stands for Retrieval-Augmented Generation.

Translate the engineering term into “daily workflow”:

Imagine you’re a strategy consultant preparing an industry brief.
You don’t:

- Sit down and type from memory.

Instead, you:

- Go to the records room (or internal knowledge base)
- Pull prior research reports
- Scan the latest few quarters’ financials
- Review news and key events
- Gather the “most relevant” materials onto your desk

Then you:

- Extract key points
- Compare and contrast
- Draw conclusions
- Write a brief for decision-makers

RAG mirrors those three steps:

- Retrieval:
  Use vector DBs, full-text search, GraphRAG, etc.,
  to pick the parts most relevant to the question from a large corpus.

- Augmentation:
  Feed those passages, graphs, and data
  as “context” to the LLM.

- Generation:
  After “reading” that context, let the LLM apply its implicit world model to:
  interpret, rewrite, reason, and summarize.

In short:

- Without RAG:
  “I know common industry templates. I’ll tell you a plausible version.”

- With RAG:
  “I reviewed your specified materials first, then used my understanding of the world to produce a tailored analysis.”

5) GraphRAG: when the task isn’t “find some passages,” but “surface the relationship network”

Vanilla RAG often retrieves “a few related documents.”
But many business questions can’t be answered by a single document.

Examples:

- “How do these three companies interact across the supply chain?”
- “Which M&A deals reshaped the industry in recent years?”
- “In our internal project history, which decision points proved pivotal?”

These are about:
- Identifying nodes (companies, projects, products, events)
- Seeing how they connect
- Then having the LLM explain “what the graph implies”

GraphRAG adds a step before RAG:

From a large document pool, it extracts:
- Names (people, companies, projects, products), time, locations
- Relationships (investment, partnership, competition, upstream/downstream, etc.)

It then builds a “graph”:
- Nodes are entities
- Edges are relations

When you query,
it doesn’t just fetch “similar passages,”
it traverses the graph to find the subgraph most relevant to your question.

Finally, it converts that subgraph back into LLM-readable text (or structured data)
for interpretation.

In effect:

- Plain RAG: from a sea of articles, pick the paragraphs you should read.
- GraphRAG: from the entire relationship graph, pick the lines you should focus on now.

6) For enterprises, where does this “consultant + records room” combo land?

With this framing, lots of use cases can be redesigned.

1. Internal knowledge-base Q&A

Scenario:
A new hire asks, “What’s our pricing strategy for this customer segment?”

Approach:
- RAG retrieves relevant content from internal docs, slides, and meeting minutes
- LLM consolidates into a response that is:
  - Clearly structured
  - Version-aware
  - With cautions and notes

Benefit:
The answer isn’t “made up from memory,” but “supported by sources plus interpretation.”

2. Investment research / industry scanning

Scenario:
You want a fast grasp of a new sector, or track its evolution over time.

Combo:
- External: RAG / GraphRAG across news, analyst reports, official sources, earnings calls
- Internal: RAG across your prior memos, models, and meeting notes
- LLM: synthesizes into:
  - Industry structure
  - Key players
  - Critical risks
  - Scenario analysis

This is sturdier than using search alone or asking an LLM alone.

3. Compliance, contracts, and policy interpretation

Scenario:
Legal or compliance teams ask, “What concrete impacts does the new regulation have on us?”

Flow:
- RAG finds relevant clauses, local interpretations, trade association commentary, internal compliance docs
- LLM:
  - Compares old vs. new requirements
  - Lists “internal processes that must change”
  - Summarizes for leadership

Key point:
The text of the rules still comes from original documents; the LLM handles interpretation and organization.

7) The most common pitfalls when designing these systems

Teams often hit similar issues when introducing RAG + LLM.

1. Expecting “the LLM will just find it on its own”

Problems:
- You never know whether it “actually saw” the right document
- You can’t control whether it used the latest version
- Errors are hard to trace to the offending source

Healthier design:
- Retrieval (RAG) is explicit and inspectable;
- Generation (LLM) is grounded and traceable.

2. Throwing raw documents into a vector DB without structuring

Plain text works, but some data should be structured:

- Timelines
- Graphs (entities/relationships)
- Tables (headers + fields)

GraphRAG shines here:
organize the “graph” first, then have the LLM interpret the graph—rather than drowning in raw text.

3. Failing to separate “who owns facts” vs. “who owns interpretation”

A simple, crucial rule:

- “Facts” must be traceable (a passage, a document, a table row)
- “Conclusions and recommendations” can be generated by the LLM via its world model

Then when something’s wrong, you can locate which fact needs updating—
instead of the whole system looking like it’s “making things up.”

Conclusion: Truly powerful AI systems look like “teams,” not “single models”

In one line:

The LLM is a well-read, reasoning, articulate consultant;
RAG, GraphRAG, and vector DBs are its records room and research assistants.

- Implicit knowledge (model weights): delivers understanding, reasoning, style, and consistency
- Explicit memory (external retrieval): delivers freshness, precision, and traceability

Together, you get an AI partner that:

- Reads documents
- Finds sources
- Explains meaning
- Communicates clearly
- And offers a clear audit trail when mistakes happen

And that’s just “how to use the model.”

Next, we’ll pivot to “how the model grows”:

- Pretraining as the “accumulation phase”: load up on world regularities, build the base
- Fine-tuning as “consolidation”: adjust style and direction, remove unnecessary noise
- RLHF as “lift-off”: express capabilities aligned with human preferences

We’ll devote a full chapter to telling “Pretrain / Fine-tune / RLHF” as a market-operator’s story.

========

This work is created by Dalishi’s AI World and released under the Creative Commons Attribution–NonCommercial–NoDerivatives 4.0 International license.

留言