RAG / Knowledge Base Engineering (Cross-Author Topic Synthesis)

When to open: You want to build a knowledge base / cannot choose between RAG vs Long Context vs NotebookLM / use RAG but get inaccurate answers / have heard of GraphRAG / LightRAG and wonder if it is worth using / want to understand why Claude Code does not use RAG. This wiki integrates perspectives on the RAG topic from 9 author wikis.

One-sentence core:RAG does not "make the model stronger"; it "automatically strengthens the prompt"(Lao D's counterintuitive judgment);Currently (2026), RAG is being replaced in 3 major directions: Long Context takes local scenarios / NotebookLM takes personal scenarios / Agentic Search takes code scenarios;Only 3 scenarios remain where RAG still has value: total document volume exceeds Long Context capacity / multi-tenant isolation / real-time indexing.

0. The "RAG / Knowledge Base" from 5 perspectives

Perspective	Source	Core positioning	One-sentence stance
5-generation architecture evolution school	Full evolution of newtype knowledge systems and Life OS / newtype · RAG Practice and Knowledge Base Evolution(Huang Yihe)	AnythingLLM slaughterhouse → Cursor+Obsidian → Claude Code AI OS → newtype OS → wiki.newtype.pro	"Your own stuff → Obsidian, other people's stuff → NotebookLM"
Counterintuitive teaching school	dontbesilent AI self-media course / 13-lesson systematic methodology §L03 (Lao D)	"RAG does not make the model stronger; it automatically strengthens the prompt"	"ima knowledge base / Yuanqi / Perplexity are all RAG at the bottom layer"
LLM interview technical school	LLM interviews / complete special topic on Retrieval-Augmented Generation(14 special topic PDFs)	R/A/G three modules + PDF parsing + table recognition + text chunking + HYDE + hard negative samples + RAG-Fusion + GraphRAG	"6 major pain points of RAG + optimization strategies"
Agentic Search school	Wang Kai's Claude Code terminal workflow methodology (complete version) + Claude Code best practices (Cal Rueb · Anthropic internal talk)(Cal Rueb)	In code scenarios, Agentic Search outperforms RAG	"Cline rejects RAG / Claude Code's built-in grep+glob is stronger than RAG MCP"
Data accuracy school	Wang Kai Agent engineering practice §3 (Wang Kai)	Self-built data MCP (FMP) replaces AI web search	"The 'numbers' for investment decisions must go through MCP, not RAG / not web search"

1. Cross-author consensus (6 points)

1.1 The three RAG modules R/A/G (industry consensus)

LLM interviews / complete special topic on Retrieval-Augmented Generation §Section 1:

Module	Function
R - Retriever	Convert external knowledge sources into vectors, and retrieve relevant documents based on the query
A - Augment	Concatenate retrieval results with the user query into a new prompt
G - Generator	The large model generates the final answer based on the augmented prompt

6-step workflow:

1. 知识库构建:文档 → 切分 chunk → embedding → 向量库
2. 用户提问:Query → embedding
3. 检索:向量相似度搜索 Top-K chunk
4. 增强:Query + 检索 chunk → 新 Prompt
5. 生成:LLM 基于新 Prompt 生成答案

Why RAG is needed(no fine-tuning / In-Context Learning):

Fine-tuning has high cost / catastrophic forgetting / long training time
In-Context Learning has limited context length
RAG = flexible + low cost + easy to update

1.2 Lao D's counterintuitive judgment: RAG does not make the model stronger (confirmed by all authors)

dontbesilent AI self-media course / 13-lesson systematic methodology §L03 The most straightforward popular science:

The most common misconception: "The knowledge base makes the large model stronger."

The truth:The model has not changed. RAG only modifies your prompt input.

工作顺序(以ima知识库为例):
1. 用户输入提问("做生意第一步...")
2. RAG 用向量检索去文档库找相关资料(找出 27 篇)
3. 把 [27 篇资料 + 用户问题] 一起灌给大模型
4. 大模型基于 28 个信息源综合输出

Why you can get good answers even with rough questions: It is not that the model has become smarter, but that the prompt has been automatically strengthened.

A knowledge base is actually an agent in disguise:

When you create an ima knowledge base, there is actually a system prompt inside (you just cannot see it)
When a question goes beyond the scope of the knowledge base, it actively refuses to answer(indicating a built-in instruction: "if the document does not contain it, refuse to answer")
Perplexity uses the same architecture —— Perplexity = knowledge base + the internet as the "knowledge source"

1.3 Document preprocessing is the most important optimization (Lao Huang consensus)

newtype · RAG Practice and Knowledge Base Evolution §2.1:

PDF → Markdown(using Marker / MinerU / Markitdown / Microsoft markitdown-mcp)
Direct use of PPT is not recommended(visual logic cannot be understood by the model) → write down the text of each page separately
Use Excel with caution(RAG handles structured data poorly)
Each section should be self-contained + have clear boundaries(ambiguous documents will definitely break RAG)

Lao Huang's 2024-03 post 36:

"If the content in the document is ambiguous, it will definitely affect the answer quality of the large model.I always try to manually revise documents into a simple form, for example, one concept or question corresponding to one answer. "

LLM interviews / complete special topic on Retrieval-Augmented Generation 5 major difficulties in PDF parsing:

Multi-column layouts (double-column academic papers)
Embedded charts (table figures in financial reports)
Character encoding anomalies (irreversible custom fonts)
Scanned image PDFs (no text layer, require OCR)
Loss of document structure information (heading hierarchy / paragraph attribution)

1.4 Model reasoning ability is the ceiling (7B is very weak in RAG scenarios)

newtype · RAG Practice and Knowledge Base Evolution §2.4 Key insight (Lao Huang 2025-05 post 1076):

Real-world testing:

Public Qwen Max: high accuracy rate
Local DeepSeek 70B / Qwen 2.5 70B: accuracy rate not high

Lao Huang's explanation:

"The comprehensive capabilities of models are worlds apart. Qwen Max's training resources and inference resources far exceed quantized versions run locally. It is much stronger in instruction following.70B open-source models easily deviate from the given context, or rely on internal knowledge rather than retrieved information. "

Conclusion:7B-14B local models are very weak in RAG scenarios. First use GPT-4 to validate the workflow, then scale down to local models.

1.5 RAG is not good at global reasoning (Lao Huang + Cline consensus)

newtype · RAG Practice and Knowledge Base Evolution §2.5:

"For the RAG technology currently used by these applications, the answer quality is poor for questions that require global understanding. When I use it personally, I only put QA and documents where local text contains complete meaning (such as notes) into the knowledge base. "

Fatal weaknesses of RAG:

Global questions(such as "what is the main idea of this article") → RAG only looks at fragments and answers inaccurately
Cross-paragraph reasoning → retrieval may miss key fragments
Tables / charts → RAG handles them poorly

Core reason Cline rejects RAG(newtype · AI Coding Practice and Tool Stack Evolution §4):

"A function,defined at line 200, called at line 300, with related dependencies at line 400. RAG can only read fragments,which breaks the logic. "

1.6 Database + model decoupling (Lao Huang Milvus + MCP)

newtype · RAG Practice and Knowledge Base Evolution §5 (key point from 2025-04 post 1083):

Approach:

Store documents in Milvus, vectorize locally
Through Milvus MCP,any AI client(Cursor / Claude Desktop / ChatWise) can access it

Why this combination is good:

"This is more flexible than previous solutions, because it separates the database from the model. The model side can be replaced freely; any model that can call MCP can be connected."

Dimension	Traditional RAG applications	Milvus + MCP
Database + model coupling	Yes	Decoupled
Model switching	Only those supported by the application	Any LLM connected via MCP
Multi-client sharing	Not possible	OK
Cost of upgrading the model	Modify the application	0

Related: OpenMemory MCP follows the same idea (see MCP Protocol Ecosystem (Cross-Author Topic Synthesis) §4.1).

2. Explicit disagreements (3 points — not smoothed over)

2.1 Disagreement 1: How to choose between RAG vs Long Context vs NotebookLM?

newtype · RAG Practice and Knowledge Base Evolution §9 Comparison of three knowledge access methods:

Method	Capacity	Suitable for
RAG	Arbitrarily large (stored in chunks)	Local precise retrieval, but weak globally
Long Context(Gemini 1M+)	1M tokens (one book)	Strong global reasoning, but cost increases with length
NotebookLM	50-300 documents / notebook	Source-Grounding, no hallucinations, balanced

Lao Huang's choice (2026-01):

Own notes < 100MB → directly Obsidian + Copilot/MCP, sufficient
Others' materials / papers → NotebookLM(free version 50 documents, premium version 300)
Truly large-scale → feed directly to Gemini Pro (1M Context)

Where RAG still has value(only 3 scenarios remain):

Total document volume exceeds the Long Context limit (several million tokens)
Multi-tenant isolation (independent RAG library per user)
Frequently updated real-time indexes

Real disagreement:Will RAG be completely replaced?:

Lao Huang's stance: already replaced in personal scenarios (Obsidian + NotebookLM end state)
LLM interview school: technical details (HYDE / hard negative samples / RAG-Fusion / GraphRAG) still have learning value
Wang Kai's stance: in data accuracy scenarios skip RAG entirely, directly connect to APIs via MCP

2.2 Disagreement 2: Is GraphRAG worth using?

newtype · RAG Practice and Knowledge Base Evolution §3 (essence of 2024-07 posts 358, 360):

Advantages of GraphRAG:Strong at handling global questions(such as "what is the main idea of this book")

Fatal drawback (2024-07):Extremely high cost

Single document + single Q&A = extremely high cost$11(using OpenAI API)
Lao Huang tried to connect local Ollama, but the embedding step always threw errors

Perspective	Stance
Lao Huang(at the time of 2024-07)	For enterprise-level in-depth research / too expensive for personal scenarios
Lao Huang(2026 end state)	Replaced by Source-Grounded NotebookLM
LLM interview school	Still needs to be learned — GraphRAG gets a standalone dedicated topic out of 14 total special topics

Judgment:

Personal use scenario → Do not use GraphRAG; use NotebookLM (free for 50 documents)
Enterprise-level deep research + no concern for cost → It is viable to implement
Learning purpose → You still need to learn the concepts

2.3 Disagreement 3: Agentic Search vs RAG (Code Scenarios)

Perspective	Stance
Cal Rueb(Claude Code Best Practices (Cal Rueb · Anthropic Internal Talk) §2)	Claude Code does not use RAG; it uses Agentic Search —— "RAG scans a book's table of contents into vectors, then finds pages by similarity. But code is not a book — code is a graph — files import each other, functions call each other. When an LLM uses grep / glob, it is like an engineer using find / ag"
Lao Huang(newtype · AI Coding Practice and Tool Stack Evolution §4)	Comparison of 4 code retrieval methods — Cline real-time search / Claude Code Agentic Search / Cursor RAG / WindSurf low-latency RAG
Cursor(Default RAG)	Classic Embeddings + RAG,high cost-performance + limited by Embedding comprehension capability

Real disagreement:Whether RAG should be used for "strongly structured documents" like code:

Anthropic stance:Do not use RAG; Agentic Search has the highest accuracy
Cursor stance:RAG has good cost-performance, accepts loss of information fragments
Lao Huang synthesis:Short notes / QA format → RAG; code / strongly structured → real-time search; cross-fragment reasoning → Long Context

3. Lao Huang's 5 Generations of Knowledge Base Architecture Evolution (Exclusive)

Full evolution of newtype knowledge system and Life OS —— 1085 posts over 26 months, 5 generations of architecture:

graph LR
    A["2024-Q1<br/>AnythingLLM 屠宰场<br/>+ Obsidian 第二大脑"] --> B["2024-Q4<br/>Cursor + Obsidian<br/>知识库"]
    B --> C["2025-H1<br/>Claude Code AI OS"]
    C --> D["2026-Q1<br/>newtype OS<br/>多 Agent 编排"]
    D --> E["2026 中<br/>wiki.newtype.pro<br/>第三层(发布层)"]

3.1 Slaughterhouse + Second Brain (2024)

Two subsystems:

External information processing / Slaughterhouse: AnythingLLM + large model.Throw in information you encounter daily, let the LLM act as a "knife" to dismember it, clearly see the "full body structure" and "valuable parts"
Note generation / Second Brain: Obsidian + various plugins.Manually filter information layer by layer, pick out what is most worth remembering, write it down in your own words

Core ritual:Must go through your own mind, type it out with your own hands. Otherwise it will never become your own knowledge.

3.2 PAFP → Funnel Directory (2024 → 2025)

PAFP(Projects / Areas / Files / Personal) = initial Obsidian directory structure
Funnel directory(Post 857, 2025-02) = Input → Processing → Output
- Input: various articles and papers you want to read closely (collection layer)
- Processing: domain-specific accumulation (digestion layer)
- Output: final deliverables, all video scripts are stored here (creation layer)

Why: PAFP is suitable for static storage;The funnel is suitable for pipelines — information flows from Input to Processing to Output, with a clear direction.

3.3 Three-Layer Linking (Lao Huang / Karpathy LLM Wiki Consensus)

Folders = first layer of linking
Tags = second layer of linking, cross-folder logic
Internal links / wikilinks = third layer of linking,knowledge graph

Elon Musk's words: "Memory is critical for biological survival, so the brain always forgets — that is a protective mechanism. If you want the brain to remember,you must build connections". "

Obsidian three-layer linking = building connections for the brain = reducing forgetting. This is the fundamental reason Lao Huang has always chosen Obsidian over Notion.

3.4 2026 Ultimate Conclusion (Post 2201, 2026-01)

After two years of iteration, on 2026-01-03 Lao Huang gave the currently most stable conclusion:

The ultimate solution for a personal knowledge base:

Your own content, put it in Obsidian

Other people's content, put it in NotebookLM

Why each:

Your own content:Must be local-first, you cannot trust any company. Markdown documents are extremely AI-friendly, and can be processed in a highly customized way using Obsidian AI plugins or Claude Code / OpenCode
Other people's content: no need to consider privacy, pursue performance. NotebookLM is the most powerful tool at present (multimodal, podcast generation, Source Grounding, Gemini access, mobile photo upload)

4. LLM Interview Technical School (Exclusive Complete Special Topic)

Complete special topic on LLM interviews / RAG Retrieval-Augmented Generation —— Full set of 14 special topic PDFs:

Topic	Key content
11. LangChain interview	LangChain RAG practical template
12. 8 memory optimization methods for multi-turn dialogue	Short-term / long-term / summary / entity / graph / vector / RAG / Hybrid
13. LangChain RAG Q&A application practice	Baidu Baike "quinoa" data + simple RAG
14. LLM + vector database document dialogue	Framework-level implementation
15. Large model RAG experience interview	Engineering practice
16. PDF parsing	5 major difficulties + 4 solutions
17. RAG layout analysis - table recognition	Rule-based / CNN / Transformer / GPT-4V
18. RAG layout analysis - text chunking	Fixed / semantic / paragraph / hierarchical
19. Large model-assisted recall (HYDE)	Hypothetical documents
20. Hard negative sample mining	Reduce hallucinations
21. RAG evaluation	RAGAS framework
22. RAG optimization strategies	Full-stack optimization
23. Key pain points of RAG	6 major categories
24. RAG-Fusion	RRF(Reciprocal Rank Fusion)
25. Graph RAG	Knowledge graph + RAG

4.1 6 Key Pain Points of RAG

Pain point	Mitigation strategy
Chunk boundaries lose semantics	overlap chunk / paragraph alignment
Retrieval misses key fragments	HYDE / RAG-Fusion / multi-query rewriting
Contradictory retrieval results	Model reasoning integration / Re-Rank
PDF parsing failure	Marker / Donut / GPT-4V
Weak table / chart RAG	Separate extraction + table recognition
Multi-turn dialogue memory drift	8 memory schemes

4.2 8 Multi-Turn Dialogue Memory Optimizations

In-session short-term(default)
Sliding window(fixed retention of the most recent N turns)
Summary-based(summarize every N turns)
Entity extraction-based(retain only entities)
Knowledge graph-based(relationship graph)
Vector memory(embed and store past conversations)
RAG memory(retrieve past content from vector store)
Hybrid(combine the above)

5. Lao D's Domestic Adaptation (Exclusive)

dontbesilent AI self-media course / 13 lessons of systematic methodology §L03 + supporting materials:

5.1 Domestic RAG Platform List

Tool	Use case	Note
ima Knowledge Base	Domestic RAG platform	Built-in RAG + personal AI avatar
Tencent Yuanqi(yuanqi.tencent.com)	Domestic agent platform	Scan code to use without login / model can be switched to R1
DeepSeek R1	Driving agents	Usable in China / international first tier

5.2 Practice: Build a Personal AI Avatar

1. 收集你的内容:推文 + 直播文稿 + 微信群问答 + 付费资料 + 写过的书
2. 上传到ima知识库(腾讯)或元气
3. 设定系统提示词:"你是 X 的 AI 分身,基于他的内容回答..."
4. 问问题 → 比直接问大模型多 90% 个性 + 风格

Lao D's own AI avatar already has over 1000 users.

5.3 Model Selection ≠ Antidote for Insufficient Prompts

"Many people, when GPT writes poorly, want to switch to Claude / Gemini. But they ignore that the core is the prompt itself. Improving prompts is more important than switching models."

6. Karpathy LLM Wiki Thinking (Lao Huang Exclusive)

Full evolution of newtype knowledge system and Life OS §10:

The "LLM Wiki" concept proposed by Karpathy in 2024-10:

"I can put some fields I am interested in (physics, biology, etc.) into applications where I have SOTA tools. I ask them questions, add relevant notes, let them organize, visualize, link, and so on."

Lao Huang's verification:This is the natural evolution direction of the Obsidian + Cursor + Claude combination — knowledge is no longer "static notes", but "a living wiki that can be conversed with, structured, invoked, and evolved".

Three-layer linking implementation:

Folders = first layer (physical organization)
Tags = second layer (logical organization)
wikilinks [[]] = third layer (relationship graph)

Users have implemented their own memory system based on Karpathy LLM Wiki(see ~/.claude/CLAUDE.md's memory design for details).

7. RAG vs Agentic Search (Comparison of 4 Retrieval Methods in Code Scenarios)

newtype · RAG Practice and Knowledge Base Evolution §6 + newtype · AI Coding Practice and Tool Stack Evolution §4:

Tool	Method	Pros / Cons
Cline	Real-time search(file content + regex)	Accurate + real-time / high Token cost
Claude Code	Agentic Search(dynamically determine content/filenames/multi-turn)	Highest accuracy / higher Token cost
Cursor	Classic Embeddings + RAG	High cost-performance / limited by Embedding comprehension
WindSurf	Same as Cursor but low latency + enterprise	Enterprise-grade / slow initial indexing

Implications for personal RAG:

For code-type, strongly structured documents → real-time search may be better than RAG
For short notes, QA format → RAG is suitable
For cross-fragment reasoning → prioritize Long Context models (Gemini 1M+)

See Claude Code Workflow (Cross-Author Topic Synthesis) §1.2 Agentic Search vs RAG.

8. Anti-Pattern Summary

❌	✅
RAG perfectionism (chasing the strongest Embedding / optimal chunk size)	First get GPT-4 + default configuration working, then downgrade and optimize
Document dumping (thousands of PDFs directly into the repository)	Curate + preprocess + convert to Markdown format
Expecting GPT-4 level results from a 7B model	Model reasoning ability is the ceiling; first validate with GPT-4
Using GraphRAG for personal scenarios	Too expensive; replace with Source-Grounded NotebookLM
Stuffing all information into the knowledge base	Only put in QA-format + self-contained + high-quality content
Expecting RAG to perform global reasoning	For global questions use Long Context or GraphRAG
Database + model coupling	Decouple with Milvus + MCP, connect any LLM
Using RAG to process code	Cline real-time search / Claude Code Agentic Search is more accurate
Not using citation verification	Always check if citations are correct; RAG errors are very hard to self-check
Treating "setting up RAG tools" as the goal	The essence of a knowledge base is knowledge itself; tools are just carriers
Expecting RAG to "make the model stronger"	RAG is "automatic prompt enhancement" (Lao D)
Putting your own content into NotebookLM	Your own content should use Obsidian local-first
Piling other people's materials into Obsidian	Other people's materials should use NotebookLM (performance first)

9. How to Choose — Knowledge Base Path by Identity

Who you are	Primary method	Starting point
Personal beginner	Lao Huang 2026 final state	Your own content → Obsidian / Other people's content → NotebookLM
AI engineer / wanting to learn technical details	LLM interview 14 special topics	R/A/G three modules + PDF parsing + table recognition + text chunking + HYDE + hard negative samples + RAG-Fusion + GraphRAG
In China without scientific internet access	Lao D ima Knowledge Base / Yuanqi / DeepSeek R1	Upload content + system prompt + personal AI avatar
Wanting multi-client shared knowledge base	Lao Huang Milvus + MCP	Database + model decoupling / any LLM access
Code scenario	Cal Rueb / Lao Huang	Agentic Search without RAG / Gemini 2.5 Pro Scans Large Codebases
Data Accuracy / Investment Scenarios	Wang Kai's Self-Built Data MCP	No RAG / No Web Search / Direct FMP API → MCP
Enterprise-grade + Cost Is Not a Concern	GraphRAG / Enterprise RAG Tools	$11 per query is acceptable
Learning / Papers / Books	NotebookLM	Source-Grounding No Hallucination / Multimodal / Mobile Camera Capture

General Startup Sequence(For any role):

Determine knowledge sources: your own vs others'
Determine integration method:RAG / Long Context / NotebookLM / Agentic Search
Go through document preprocessing: PDF → Markdown / Each section self-contained
Validate with flagship models GPT-4 / Claude / Gemini(7B local models are weak at RAG)
Default chunk size 1000 tokens + overlap
Citations must be reviewed(RAG errors are very hard to self-check)
After stabilization → decouple the database(Milvus + MCP)
Further step → structure knowledge(Karpathy LLM Wiki 3-layer association)

10. Reference List

Primary original author wikis:

newtype · RAG Practice and Knowledge Base Evolution —— Lao Huang's complete RAG evolution / 5 key points / GraphRAG / Cohere Command R+ / Milvus+MCP / Comparison of 3 knowledge integration approaches
newtype Full Evolution of Knowledge Systems and Life OS —— 1085 posts / 5 generations of architecture evolution / PAFP→funnel / three-layer association / Karpathy LLM Wiki / 2026 end state
newtype · MCP Protocol and Ecosystem in Practice —— Milvus / OpenMemory / NotebookLM MCP / context7
newtype · AI Tool Stack Practice and Evolution —— Reviews of AnythingLLM / Open WebUI / RAGFlow / NotebookLM tools
newtype Miscellaneous Practices: llms.txt Protocol + CrewAI Obsidian Project + n8n MCP Workflow —— CrewAI Obsidian project / 3 Agents (Research+Edit+Notetake)
LLM Interviews / Complete Special Topic on Retrieval-Augmented Generation (RAG)—— 14 RAG special-topic PDFs / R/A/G / PDF parsing / tables / chunking / HYDE / hard negative samples / 6 major pain points / GraphRAG / RAG-Fusion / 8 memory optimization approaches
dontbesilent AI Self-Media Course / 13-Lesson Systematic Methodology §L03 —— "RAG does not make the model stronger; it strengthens the prompt" / ima knowledge base / Yuanqi / same origin as Perplexity / personal AI avatar
Wang Kai's Claude Code Terminal Workflow Methodology (Complete Version) §1.2 —— Agentic Search vs RAG
Claude Code Best Practices (Cal Rueb · Anthropic Internal Talk) §2 —— Cal Rueb Agentic Search / grep+glob

History

2026-05-18:Topic-dimension wiki first release. Synthesized and rewritten from 9 author wikis. Explicitly notes 3 points of disagreement: RAG vs Long Context vs NotebookLM / GraphRAG economics / Agentic Search vs RAG (code scenarios)