我的征尘是星辰大海。。。
The dirt and dust from my pilgrimage forms oceans of stars...
-------当记忆的篇章变得零碎，当追忆的图片变得模糊，我们只能求助于数字存储的永恒的回忆
作者:黄教授
手机视频列表
这个博主做了一年的RAG的知识工程他究竟懊悔的是什么
视频
音频
原始脚本
我看到一个博主做 RAG，懊悔浪费这一年，他究竟后悔的是什么？这位博主说，做 RAG 这一年最后悔的就是上知识图谱。 我感觉不是知识图谱技术不行，是你们把知识图谱做成了垃圾，不是吗？我就针对圈内这一篇，做 RAG 这一年。 最后悔的就是上知识图谱的说法，发表我的评论，看看有多少人同意。 作者痛陈三元组索引毁掉上下文语义，看似技术踩坑，实则戳穿了整个行业自欺欺人的遮羞布。 花几千万砸出来的高端知识图谱，连人类语言最基础的结构都装不下。 我们先把核心问题扒干净，所谓工业级知识图谱，底层居然还是半个世纪前的扁平三元组，投实体、关系、尾实体。 A 是药物，B 是疾病，关系是有效、完事。 依据2025年实验中，A 药对60岁以上无心脏病史的 B 病患者有效，硬生生被砍成 A 药有效、B 病。 条件呢？范围呢？前提呢？推理链呢？全扔了。 更讽 次的是，这套结构根本不是技术做不到，纯粹是偷懒守旧、利益固化。 50年前的 LISP、Prolog 就支持递归嵌套，XML、JSON 天然就是层次结构。 人类语言随便一句长难句，主句套从句，从句欠主谓宾，自带递归之实数。 连普通人都能轻松理解的复杂逻辑，今天花千万级预算打造的智能基础设施，居然装不下、存不了、索引不动。 这就是当下 L A G 加知识图谱最荒诞的现实。 大模型明明能读懂整段文本，吃透上下文，完成多跳推理，拥有接近人类的语义理解能力。 你们却非要把完整鲜活带逻辑的知识切碎、压扁、阉割成最简陋的碎片，再喂给大模型。 相当于给超级大脑喂17时代的粗粮，然后抱怨 AI 不够智能。 你们管这叫知识工程，这叫知识破坏。 你们管这叫赋能大模型，这叫拖后腿、锁上限、自残式落地。 自然语言几千年就进化出的递归、嵌套、条件、上下文，50年前的软件就能实现的层次结构，今天大模型已经炉火纯青的语义能力，偏偏被一群守着老旧标准、抱着存量利益的人困在扁平三元组的死胡同里，不敢改、不愿改、不能改，美其名曰工业级稳定通用，本质就是固步自封，敷衍了事。 花几千万做出来的东西连自然语言都不如，号称做知识智能，连知识最核心的上下文与逻辑都存不住。 用最落后的结构去绑定最先进的大模型，用石器时代的方案去服务人工智能时代的需求。 这不是技术困境，这是行业耻辱。 那位作者说，最后悔上知识图谱，悔的不是知识图谱这个方向，而是被一套过时、简陋、阉割知识的结构绑架。 毁的是明明有更好的路，却被旧体系、旧思维 就利益堵死。 我们今天把话放着，谁还在强行把文本压扁成扁平三元组，谁就是在给大模型喂垃圾。 谁还在鼓吹老旧知识图谱是 RAZA 最优解，谁就是在装睡骗钱。 知识不该是孤立的碎片，智能不该被简陋的结构锁死。 能唤醒一个是一个，能骂醒一个是一个。 别再用石器时代的思维，糟蹋工业级的人工智能。
修正脚本
我看到一个博主做 RAG，懊悔浪费这一年，他究竟后悔的是什么？这位博主说，做 RAG 这一年最后悔的就是上知识图谱。 我感觉不是知识图谱技术不行，是你们把知识图谱做成了垃圾，不是吗？我就针对圈内这一篇，做 RAG 这一年。 最后悔的就是上知识图谱的说法，发表我的评论，看看有多少人同意。 作者痛陈三元组索引毁掉上下文语义，看似技术踩坑，实则戳穿了整个行业自欺欺人的遮羞布。 花几千万砸出来的高端知识图谱，连人类语言最基础的结构都装不下。 我们先把核心问题扒干净，所谓工业级知识图谱，底层居然还是半个世纪前的扁平三元组，头实体、关系、尾实体。 A 是药物，B 是疾病，关系是有效、完事。 依据2025年实验中，A 药对60岁以上无心脏病史的 B 病患者有效，硬生生被砍成 A 药有效、B 病。 条件呢？范围呢？前提呢？推理链呢？全扔了。 更讽刺的是，这套结构根本不是技术做不到，纯粹是偷懒守旧、利益固化。 50年前的 LISP、Prolog 就支持递归嵌套，XML、JSON 天然就是层次结构。 人类语言随便一句长难句，主句套从句，从句嵌主谓宾，自带递归之属性。 连普通人都能轻松理解的复杂逻辑，今天花千万级预算打造的智能基础设施，居然装不下、存不了、索引不动。 这就是当下 RAG 加知识图谱最荒诞的现实。 大模型明明能读懂整段文本，吃透上下文，完成多跳推理，拥有接近人类的语义理解能力。 你们却非要把完整鲜活带逻辑的知识切碎、压扁、阉割成最简陋的碎片，再喂给大模型。 相当于给超级大脑喂石器时代的粗粮，然后抱怨 AI 不够智能。 你们管这叫知识工程，这叫知识破坏。 你们管这叫赋能大模型，这叫拖后腿、锁上限、自残式落地。 自然语言几千年就进化出的递归、嵌套、条件、上下文，50年前的软件就能实现的层次结构，今天大模型已经炉火纯青的语义能力，偏偏被一群守着老旧标准、抱着存量利益的人困在扁平三元组的死胡同里，不敢改、不愿改、不能改，美其名曰工业级稳定通用，本质就是固步自封，敷衍了事。 花几千万做出来的东西连自然语言都不如，号称做知识智能，连知识最核心的上下文与逻辑都存不住。 用最落后的结构去绑定最先进的大模型，用石器时代的方案去服务人工智能时代的需求。 这不是技术困境，这是行业耻辱。 那位作者说，最后悔上知识图谱，悔的不是知识图谱这个方向，而是被一套过时、简陋、阉割知识的结构绑架。 毁的是明明有更好的路，却被旧体系、旧思维、旧利益堵死。 我们今天把话放着，谁还在强行把文本压扁成扁平三元组，谁就是在给大模型喂垃圾。 谁还在鼓吹老旧知识图谱是 RAG 最优解，谁就是在装睡骗钱。 知识不该是孤立的碎片，智能不该被简陋的结构锁死。 能唤醒一个是一个，能骂醒一个是一个。 别再用石器时代的思维，糟蹋工业级的人工智能。
英文翻译
I saw a blogger who worked on RAG regret wasting the past year. What exactly is he regretting? The blogger said that the biggest regret of the past year working on RAG was adopting knowledge graphs. I feel it's not that knowledge graph technology is flawed, but that you've turned knowledge graphs into garbage, haven't you? I'm addressing this specific article in the industry about working on RAG for a year. The claim that the biggest regret was adopting knowledge graphs—I'm sharing my commentary on this to see how many people agree. The author laments that triple indexing destroys contextual semantics in sentences. It looks like a technical pitfall, but it actually exposes the self-deceptive fig leaf of the entire industry. Spending tens of millions to build a high-end knowledge graph that can't even accommodate the most basic structure of human language. Let's strip down the core problem first: so-called industrial-grade knowledge graphs are still built on the flat triple structure from half a century ago—head entity, relation, tail entity. Example: A is a drug, B is a disease, the relation is "effective"—done. According to a 2025 experiment, Drug A is effective on patients over 60 with no history of heart disease for Disease B—this gets brutally reduced to "Drug A effective, Disease B." What about the conditions? The scope? The prerequisites? The reasoning chain? All thrown away. More ironically, this structural failure isn't because the technology can't do better—it's pure laziness, conservatism, and entrenched interests. Fifty years ago, LISP and Prolog already supported recursive nesting; XML and JSON are inherently hierarchical. Any long, complex sentence in human language—main clause containing subordinate clauses, subordinate clauses embedding subject-verb-object structures—has an inherent recursive property. Even ordinary people can easily grasp complex logic, yet today's intelligent infrastructure, built with budgets in the tens of millions, can't contain, store, or index it. This is the most absurd reality of current RAG plus knowledge graphs. Large language models can clearly read entire texts, understand context, perform multi-hop reasoning, and possess near-human semantic comprehension. But you insist on chopping, flattening, and castrating complete, vibrant, logical knowledge into the simplest fragments before feeding it to the LLMs. It's like feeding a superbrain Stone Age gruel, then complaining that AI isn't smart enough. You call this knowledge engineering—this is knowledge destruction. You call this empowering large models—this is holding them back, capping their potential, self-sabotaging deployment. Natural language evolved recursion, nesting, conditions, and context over thousands of years. Hierarchical structures were achievable with software from 50 years ago. Today's large models already possess mature semantic capabilities. Yet you—stuck with outdated standards and clinging to existing interests—trap everything in the dead end of flat triples. You dare not change, are unwilling to change, cannot change, all under the banner of "industrial-grade stability and universality." In essence, it's nothing but self-imposed stagnation and perfunctory work. Spending tens of millions to produce something that can't even match natural language—calling this knowledge intelligence while failing to store the most core elements: context and logic. Binding the most advanced large models with the most backward structures. Using Stone Age solutions to serve the demands of the AI era. This is not a technological dilemma—it's an industry disgrace. The author said the biggest regret was adopting knowledge graphs. The regret isn't about the direction of knowledge graphs themselves, but about being shackled by an outdated, simplistic, knowledge-castrating structure. The tragedy is that there are clearly better paths, but they are blocked by old systems, old thinking, and old interests. Let's put it plainly: whoever still forcibly flattens text into flat triples is feeding garbage to large models. Whoever still touts old-fashioned knowledge graphs as the optimal solution for RAG is either sleeping and scamming for money. Knowledge shouldn't be isolated fragments. Intelligence shouldn't be locked down by crude structures. Wake up one person if we can. Scold one person awake if we can. Stop using Stone Age thinking to ruin industrial-grade artificial intelligence.
back to top