【AI工具】CodeGraph/Graphify/GitNe

【AI工具】CodeGraph/Graphify/GitNexus 三大开源代码图谱横评 · 告别API烧钱

发布于 2026-08-20 · 时长 11:43

字幕摘录

时间	英文	中文
0:00	Your coding agent just spent 40 tool calls answering one question about the VS Code repo.	你的编码代理刚刚花了40个工具电话回答一个关于VS代码回波的问题
0:05	One question.	一个问题。
0:06	A dollar forty-one.	一钱四十一.
0:08	And 17 files read before it even started thinking.	在它开始思考之前读了17份文件
0:11	Same question, same model.	同样的问题,同样的模型。
0:12	With a code graph in front of it.	前面有一个代码图
0:14	Two calls.	两通电话
0:15	Thirty-six cents.	三十六钱.
0:16	Three open source tools sell you that trade.	3个开源工具卖你的交易。
0:19	Graphify.	图形化。

展开字幕全文（225 条）

序号	英文	中文
1	Your coding agent just spent 40 tool calls answering one question about the VS Code repo.	你的编码代理刚刚花了40个工具电话回答一个关于VS代码回波的问题
2	One question.	一个问题。
3	A dollar forty-one.	一钱四十一.
4	And 17 files read before it even started thinking.	在它开始思考之前读了17份文件
5	Same question, same model.	同样的问题,同样的模型。
6	With a code graph in front of it.	前面有一个代码图
7	Two calls.	两通电话
8	Thirty-six cents.	三十六钱.
9	Three open source tools sell you that trade.	3个开源工具卖你的交易。
10	Graphify.	图形化。
11	GitNexus.	吉特内克斯.
12	代码Graph。
13	200,000 GitHub stars between them.	他们之间有20万颗巨星
14	And most of it arrived this year.	今年大部分时间都到了
15	Star count is the worst possible way to choose one of them.	星数是最糟糕的选择其中之一的方法.
16	You choose by answering three things.	你通过回答三件事来选择
17	What gets indexed?	什么被索引了?
18	What it costs?	多少钱?
19	What it locks you into?	它让你陷入什么困境?
20	One of these three bands commercial use in its license.	这三个乐队之一在其执照中商业使用。
21	Forty-four thousand stars you can't ship at work.	四万四千颗星你不能在工作上飞船
22	Another one's readme describes a product its own package file says it doesn't ship.	另一个是readme描述一种产品,它自己的包文件说它不会飞船.
23	Start with what they share, because it's more than you'd guess.	先从他们分享的开始, 因为它比你想象的还要多。
24	All three parse your code with Treesitter.	三个都把代码和树卫分析出来
25	No model in that step.	那一步没有模特儿
26	No embeddings.	没有嵌入。
27	Nothing leaving the machine.	没有东西离开机器。
28	The graph that falls out is the same shape too.	掉下来的图也是同样的形状.
29	Functions.	职能。
30	Classes and files as nodes.	类和文件作为节点.
31	Calls, imports, and inheritance as edges.	要求,进口,和继承作为边缘。
32	Everything after that is where they disagree.	之后的一切都是他们不同意的地方。
33	And that's where you pick.	这就是你选的地方。
34	CodeGraph first, because it's the only one of the three that publishes a benchmark you	代码Graph第一,因为它是唯一三个发布基准你
35	can rerun yourself.	可以重新运行自己。
36	Sixty-two thousand stars, MIT licensed, six months old.	6万2千颗恒星,麻省理工学院特许,6个月大.
37	Colby McHenry created the repo in January.	科尔比·麦克亨利于1月创建了雷波.
38	And in May it hit number two on GitHub trending with 2,400 stars in a single day.	而在5月,它在GitHub上以一天2400颗恒星的热潮位居第二.
39	Open its GitHub page and the language badge says C.	打开它的GitHub页面,语言徽章写着C.
40	That is technically true and completely misleading.	这在技术上是真实的,而且完全是误导性的。
41	Repo持有21 C文件.
42	Two of them are generated Treesitter parser tables, Scala and Kotlin.	其中两个是生成的树人解剖表,斯卡拉和科特林.
43	And those two alone are 58 megabytes.	而这两个单是58兆字节。
44	人类写的代码是341 TypeScript文件和26 Rust文件.
45	Those Rust files are the whole story of version 1.5 shipped on the 21st of July.	这些Rust文件是7月21日发货的1.5版的全部故事.
46	20种CodeGraph的34种语言,
47	per file.	每个文件。
48	其余14个掉回Web Assembly Treeser.
49	And the project's claim is that both paths produce byte-identical graphs.	并且项目的索赔是,两条路径都产生字节相同的图表.
50	输出在 SQLite 文件, .codegraph slash codegraph.db 中登陆 FTS 5 全文搜索
51	on top.	在顶部。
52	No vector store, no embedding model, no API key, and nothing leaves your machine.	没有向量存储器,没有嵌入模型,没有API密钥,没有任何东西离开你的机器.
53	The whole index is a file you can delete when you get bored of it.	整个索引是一个在厌倦时可以删除的文件.
54	Then it watches.	然后它看着。
55	FSEvents on macOS, Enotify on Linux, Read Directory changes W on Windows. 互联网档案馆的存檔,存档日期2013-12-21.
56	Save a file, and the graph catches up in about three-tenths of a second on a 4,000 file project.	保存一个文件,图表在一个4000个文件项目上追上大约三十分之一秒.
57	Four-tenths on the 27,000 file Swift compiler.	27,000文件Swift编译器上的4分之10.
58	It never rescans the tree.	它从不重新扫描树。
59	And when it hasn't caught up, it says so out loud.	当它没有赶上, 它说得那么大声。
60	The tool response gets a warning at the top naming the stale file, telling the agent to	工具响应在顶端命名 stale 文件时得到警告, 告诉代理
61	read that one directly and trust the rest.	读一个,直接相信其余的。
62	A tool that admits when it's behind is rarer than it should be.	一种在落后时才承认的工具比应该承认的工具更罕见.
63	There's also a command I did not expect to like this much.	还有我没想到会这么喜欢的命令
64	Pipe a git diff into code graph affected, and it hands back the test files your change	将一个 git diff 插入到受影响的代码图中, 它把测试文件递回您的更改
65	can actually reach.	可以实际到达。
66	Four lines of shell, and your test suite stops running everything on every commit.	四行的壳, 你的测试套房停止运行每一次承诺。
67	Here is the benchmark.	以下是基准。
68	Seven real repositories, seven languages, one architecture question each, median of	7个真正的寄存器、7种语言、1个建筑问题、
69	four runs, revalidated on the 21st of July against Claude Opus 4.8, VS code, two tool	4次运行,7月21日对克劳德·奥普斯4.8,VS代码,两个工具进行重新验证
70	calls instead of 40, Django, two instead of 29, Tokyo, three instead of 57.	电话40,强哥,2,29,东京,3,57
71	File reads zero on all seven.	所有7个文件都读为零。
72	In money, Tokyo dropped from $3.04 to 44 cents.	在金钱上,东京从3.04美元下降到44美分.
73	Excali draw from $1.81 to 40, averaged over the seven repos, 89% fewer tool calls, 69%	Excali从1.81美元到40美元,平均为7次重置,减少89%,69%
74	fewer tokens, 60% cheaper.	代币少,便宜60%.
75	But read the footnotes because they left the losses in.	但读一下脚注,因为他们留下了损失。
76	On Excali draw, the graph run took 36 seconds and plain grep took 23.	在Excali画上,图跑用了36秒,平原的grep用了23秒.
77	On OKHTP, it cost three cents more.	在OKHTP上,它花费三美分多.
78	Their own explanation.	他们自己的解释。
79	On a small repo opus, grep's fast enough to win the clock while spending five to ten	在一次小的回波行动上,grep的速度足以赢得时钟同时花费5到10美元
80	times the tokens.	倍数的标志。
81	The number that convinced me is the Linux kernel.	说服我的数字是Linux内核
82	70,000 files, two million symbols, 6.4 million relationships.	7万个档案,200万个符号,640万个关系.
83	Indexed to completion in under 12 minutes on a two-core machine with six gigabytes of	指数在12分钟内完成拥有6千兆字节
84	memory, where a memory first design runs out before it reaches 1%.	内存,一个内存的第一个设计在达到1%之前就已耗尽。
85	What you give up is everything that isn't code.	你放弃的不是密码
86	No PDFs, no design docs, no notes.	没有PDF,没有设计文件,没有注释.
87	And its own instructions warn that if your agent hands exploration to a sub-agent, that	它自己的指示警告,如果你的代理人向副代理人探索,
88	sub-agent reads files anyway and code graph becomes pure overhead.	子代理读取文件, 代码图会变成纯的间接费用。
89	It only helps when it is asked directly.	只有在被直接询问时才有用。
90	Graphify is the one you have already seen on your timeline, 95,600 stars.	你在时间线上已经看到了95 600颗恒星
91	The repository was created on the 3rd of April this year.	该寄存器于今年4月3日创建.
92	That is 113 days to 95,000 stars.	这是113天到95,000颗恒星。
93	A rate almost nothing sustains.	利率几乎没有维持。
94	It is a Y Combinator company, summer 26 batch, with a readme translated into 35 languages	是一家Y联合公司夏天26批读数被翻译成35种语言
95	and an enterprise waitlist behind it.	并有一个企业在它的后面等待。
96	A funded product with an open source front door, which is worth knowing before you build	开源前门的融资产品值得你先知道
97	on it.	交给我
98	The pitch is a different shape from code graphs.	投球是不同于代码图的形状.
99	You type slash graphify in your assistant and it maps the project.	您在助手中输入斜线图形化, 它会映射项目。
100	Not only the code, markdown, PDFs, spreadsheets, screenshots, and video all land in the same	不仅仅是代码、标记下、PDF、电子表格、截图和录像
101	graph as your functions.	图表为您的函数。
102	That splits the tool in half.	这把工具分成两半
103	Code goes through Treesitter locally, deterministic, zero tokens, nothing leaving your machine,	密码通过本地的植树工, 确定性,零符号, 没有什么离开你的机器,
104	and a code only repo needs no API key at all.	而一个代码只需要 repo 不需要 API 密钥。
105	Everything else goes to a model.	其他的都交给模特儿
106	Video and audio get transcribed locally with Faster Whisper first.	视频和音频在本地进行转录,先使用更快的Whisper.
107	36 Treesitter grammars, 79 file extensions, and then a long tail nobody else bothers with.	36树人语法,79个文件扩展,然后是一条长尾巴,没人会打扰.
108	Salesforce Apex, Terraform, Package Maniform, MCP 配置文件.
109	Hand it a connection string and it will read a live Postgres schema straight into the graph.	把它交给一个连接字符串, 它会直接读出一个活的 Postgres 的图。
110	Every edge carries a tag.	每一个边缘都有标记
111	Extracted means it is written in the source.	被提取表示它是写在来源。
112	Inferred means graphify resolved it.	推论意味着用图解它。
113	Ambiguous means it is not sure.	含糊不清意味着不确定。
114	You can point at any connection and know whether it was read or guessed, which is the thing	你可以指向任何连接,并知道它是读的还是猜到的,这是事情
115	I want most from a tool like this.	我最想从这样的工具。
116	Three files come out.	三个文件出来。
117	An interactive HTML graph.	交互式 HTML 图表。
118	A JSON graph you can query weeks later without re-reading anything.	一个JSON图,你可以在几周后查询,而不重新阅读任何东西.
119	And a markdown report listing what it calls godnodes, the highest degree concepts that	还有一份降级报告列出了所谓的Godnodes 最高学位的概念
120	everything in the project routes through.	所有在项目路线通过。
121	Now the part that does not come up.	现在,没有出现的部分。
122	The old readme's headline number was 71 and a half times fewer tokens per query, measured	旧读器的头条号码是71倍半减少的每个查询,测量
123	on a 52 file corpus.	在一个52个档案里
124	Search the current readme for that number.	搜索当前读取的号码。
125	It is not there.	它不在这里。
126	Not revised down.	未修订下来。
127	Removed.	已删除。
128	What replaced it is a real benchmark table.	取而代之的是真正的基准表。
129	Locomo, 300 问.
130	Recall at 10 of .497 against Mem0 at .048 and SuperMemory at .149.	回顾497年10月10日对梅姆0战绩为0.048,超级记忆战绩为149.
131	A tenfold gap on retrieval, scored by a judge with 90.6% agreement against a second judge.	在检索方面有10倍的差距,一名法官以90.6%的协议对第二名法官进行评分。
132	And then the row most companies would have deleted.	然后大多数公司都会删除这行。
133	On question answering accuracy, graphify scores 45.3% and SuperMemory scores 49.7.	关于答题准确性,图化得分45.3%,Super Memory得分49.7.
134	It loses that one.	它失去了那个。
135	On long memeval, it ties with plain dense retrieval.	在长的memeval上,它与纯密的检索有关.
136	Publishing your own losing row buys a lot of trust from me, which tells you what graphify	出版你输掉的专栏赢得了我很大的信任这告诉你什么是图解
137	really is.	痷
138	A memory layer that happens to be very good at code, benchmarked against memory systems	一个恰巧非常擅长代码的内存层,以内存系统为基准
139	instead of code indexers.	而不是代码索引器。
140	If your answer lives half in the repo and half in a folder of PDFs, this is the only	如果你的答案一半和一半在 PDF 文件夹中, 这是唯一的
141	one of the three that even tries.	其中一个甚至尝试。
142	GitNexus is the oldest of the three, created in August 2025, five months before CodeGraph	GitNexus是三部曲中最古老的一部,创作于2025年8月,比CodeGraph早5个月.
143	and eight before graphify.	和8个之前的图解。
144	44,600 stars, and by a wide margin the largest tool surface here.	44600颗恒星, 和宽度最大的工具表面在这里。
145	17个MCP工具,其中CodeGraph刻意将1号舰载.
146	Impact for blast radius, trace for the shortest path between two symbols.	对爆炸半径的冲击追踪两个符号之间最短的路径
147	Detect changes, which maps a git diff onto the execution flows it touches.	检测变化, 将 git diff 映射到它碰到的执行流上。
148	Rename, which coordinates edits across every file that references a symbol.	重命名,它协调每个引用符号的文件的编辑。
149	Underneath, 15 pipeline phases, 44 node types, 21 relationship types.	地下有15个管道阶段,44个节点类型,21个关系类型.
150	It runs Lydon clustering to find communities in your code, then traces execution flows	它运行 Lydon 集群以在您的代码中找到社区,然后追踪执行流
151	from entry points through the call chain, and stores those as process nodes you can	从通过呼叫链的切入点,并将这些存储为进程节点
152	query by name.	按名称查询。
153	And it is the only one of the three you query in Cypher.	这是你唯一一个在Cypher查询的
154	Match a community labeled authentication.	匹配社区标签认证。
155	Match everything calling into it.	配合一切调用它。
156	Filter to edges above 0.8 confidence.	过滤到0.8置信度以上的边缘。
157	If you have written a Neo4j query, you already know how to drive this.	如果您已经写了 Neo4j 查询, 您已经知道如何驱动。
158	Now the claim on the tin.	现在对锡的要求。
159	README说,网页界面完全运行在 Web Assembly, Treesitter Was, Database中
160	WAS, in browser embeddings, and that no code is uploaded to any server.	WAS,在浏览器嵌入中,没有代码被上传到任何服务器.
161	So open the web apps package file.	所以打开网络应用软件包文件.
162	没有WebTreesitter,没有WASM数据库,也没有变压器.js.
163	The only file in its services directory is an HTTP client.	其服务目录中唯一的文件是HTTP客户端.
164	Scan all 6,500 files in the repository and every WASM reference sits inside the Node	扫描寄存器中的全部 6,500 文件, 每个 WASM 参考文件都坐落在节点内
165	command line tool.	命令行工具。
166	The live site agrees with the code rather than the README.	现场站点同意代码,而不是README.
167	Load the GitNexus page today and it is a three-step wizard telling you to run npx git nexus serve	今天装入 GitNexus 页面, 它是一个三步向导, 告诉你运行 npx git 连线服务
168	and wait on port 4747.	在4747号港口等候
169	It is a WebGL frontend for a local node server.	它是本地节点服务器的 WebGL 前端。
170	That is Drift, not Deception.	这是漂流,不是欺骗。
171	The architecture moved and the documentation did not follow.	该架构被移动,文档没有随之而来.
172	But if you picked this tool because it ran in a browser tab with nothing installed, you	但如果你选中了这个工具,因为它运行在浏览器标签中,没有安装任何设备,那么你就
173	picked it for something that is not shipping anymore.	选它是为了不再运输的东西
174	Then there is the license.	还有执照
175	GitHub's detector prints no assertion, so almost nobody opens the file.	GitHub的探测器没有打印任何断言,所以几乎没有人打开文件.
176	Open it, polyform non-commercial 1.0.	打开它,多形式非商业的1.0。
177	Any non-commercial purpose is permitted.	允许任何非商业性目的。
178	Commercial use is not one of them.	商业用途不是其中之一。
179	Schools, charities, research labs, government, your weekend project, all fine.	学校慈善机构研究实验室政府周末计划都很好
180	Your employer's monorepo, not fine, unless somebody buys a commercial license.	你老板的单曲,不是罚款, 除非有人购买商业执照。
181	放在CodeGraph的麻省理工学院和Graphify的Apache 2.0旁边这是最大的区别
182	on this list.	名单。
183	One more thing.	还有一件事
184	Git nexus publishes no benchmark at all.	Git 連結完全没有发布基准.
185	No index time, no node counts, no token savings, no latency.	没有指数时间,节点计数,没有象征性的储蓄,没有延迟.
186	It ships a SFUE bench harness with three modes and eight models wired up and zero results.	它搭载一个SFUE长椅吊带,有三种模式和八种模式连接,结果为零.
187	The harness states a hypothesis.	绳子是假设
188	The repository never answers it.	寄存器从不回答.
189	Head to head then.	头到脑后。
190	What gets indexed?	什么被索引了?
191	CodeGraph, 34 languages, code only, plus route extraction for 17 web frameworks.	CodeGraph,34种语言,只有代码,加上17个网络框架的路由提取.
192	Graphify, 36 grammars plus docs, PDFs, images, and video.	图表化,36个语法加上文件、PDF、图像和视频。
193	Git nexus, 14 languages, code plus markdown sections, what it costs to run.	Git连线,14种语言,代码加减码段,运行成本.
194	CodeGraph indexes for zero tokens, but anonymous telemetry is on by default and you opt out.	代码Graph指数表示零符号,但匿名遥测默认进行,你选择退出.
195	Graphify is zero tokens for code and model calls for everything else, with no telemetry	图形化是零符号代码和模型调用其他一切,没有遥测
196	at all.	一点也不
197	Git nexus index is free, but its wiki generator and its chat agent both want an API key.	Git连线索引是免费的,但其维基生成器和聊天代理都想要API密钥.
198	What it locks you into.	它让你陷入什么境地
199	麻省理工学院,Apache 2.0, Polyform 非商业性.
200	If you work at a company, that third one already made your decision for you and what each one	如果你在一家公司工作那第三家已经为你做了决定
201	will prove to you.	将证明给你。
202	CodeGraph publishes seven repositories with clone commands and per run dollar costs.	CodeGraph发布7个寄存器,包含克隆指令和每运行美元成本.
203	Graphify publishes third party memory benchmarks with a judge agreement score.	图表公布第三方记忆基准,并附有法官协议分数。
204	Git link发布一个假说.
205	So here is how I would pick.	所以我会选择这样
206	Large code base, agent burning money on discovery.	大型密码基地特工在发现后会烧钱
207	拿着CodeGraph
208	It is the only one that shows you receipts, it is MIT, and the Linux kernel result means	这是唯一一个显示你收据,它是麻省理工学院, Linux 内核结果意味着
209	it will not fall over on your monorepo.	它不会掉到你的单簧管上
210	Context that is not only code, where the answer lives half in the repo and half in a folder	不只是代码, 答案在回存中半, 在文件夹中半
211	PDF和截图。
212	Take Graphify.	以图法.
213	Nothing else here attempts that, and the code path still runs local and free.	这里没有其他的尝试, 代码路径仍然运行本地和免费。
214	Deepest query surface, and you are not shipping commercially.	最深的查询表面, 而你不是商业运输。
215	Take Git nexus.	取吉特脉络.
216	17 tools, cipher, taint analysis, execution flows.	17个工具,密码,纹理分析,执行流.
217	It is the researcher's tool of the three.	这是三者研究者的工具。
218	And the license says the same thing.	且牌照说同.
219	One thing worth saying out loud.	有件事值得大声说
220	These three are not really fighting over one slot.	这三人其实不是在一个档次上争吵.
221	One index is code fast, one index is everything, one query is deepest.	一个指数是代码快速,一个指数是一切,一个查询是最深的.
222	Running two of them at once is a perfectly sane answer.	一次跑两个是完全正常的答案
223	Pick the one that matches the question you keep asking your agent, then open the license	选一个符合你一直问的问题,然后打开许可证
224	file before you commit to it.	在你承诺之前先存档。
225	That is the video.	这是视频。

该视频共有字幕 225 条。解锁更多字幕为会员功能，请移动到价格

下载安装