"AI-readable Markdown" — can't we just say "Markdown"? I'm serious about this, why are we focussing on making things accessible to AI when they should just be machine-readable and accessible to human beings in the first place? No need to taint this by bringing AI into it.
> why are we focussing on making things accessible to AI
Because that’s the authors actual goal? To take a web page that looks fine to human eyes but is unintuitively not accessible to AI. That’s genuinely useful and valuable.
Sure it’s no different than converting it to markdown for human eyes. But it’s important to be clear about not just WHAT but also WHY.
To achieve its goal I would be worried that it dropped the verbatim function signature. Claude still figured it out, but for more obscure stuff that could be an issue.
Following up — I just pushed a fix for this. This latest version significantly improves how references like protocol conformances and default implementations are rendered.
Nice! Do you think it could be adapted to other docs sites?
I made a small clone of the tutorials section (https://clone-swiftui-tutorial.vercel.app/) where the content is already Markdown (and use codehike to turn the markdown into a rich UI). This made me realize that codehike is AI-friendly, in the sense that even for non-linear UIs the original content is still AI-readable Markdown.
How to reliably HTML to MD for any page on the internet?
I remember struggling with this in the past
How hard would it be to build an MCP that's basically a proxy for web search except it always tries to build the markdown version of the web pages instead of passing HTML?
Basically Sosumi.ai but instead of working on only for Apple docs it works for any web page (including every doc on the internet)
I think HTML -> Markdown is a bit of a red herring.
In many cases, a Markdown distillation of HTML can improve the signal-to-noise ratio — especially for sites mired in <div> tag soup (intentionally or not). But that's an optimization for token efficiency; LLMs can usually figure things out.
The motivation behind Sosumi is better understood as a matter of accessibility. The way AI assistants typically fetch content from the web precluded them from getting any useful information from developer.apple.com.
You could start to solve the generalized problem with an MCP that 1) used a headless browser to access content for sites that require JS, and 2) used sampling (i.e. having a tool use the host LLM) to summarize / distill HTML.
But stripping complex formats like html & pdf down to simple markdown is a hard problem. It's nearly impossible to infer what the rendered page looks like by looking at the raw html / pdf code. https://github.com/mozilla/readability helps but it often breaks down over unconventional div structures. I heard the state of the art solution is using multimodal LLM OCR to really look at the rendered page and rewrite the thing in markdown.
Which makes me wonder: how did OpenAI make their model read pdf, docx and images at all?
There are APIs such as Jina AI's reader API that do this pretty well. It doesn't produce output as clean as Sosumi for Apple docs, but it's free and does a decent job.
Hm, I would have extracted the markdown from the Swift source code. That’s what Apple uses to generate their pages, using https://www.swift.org/documentation/docc/.
This is only possible for some of Apple‘s open source Swift code, including the Swift standard library. This is not the case for hundreds of other SDK frameworks, such as SwiftUI.
Even for those open source projects, there is still some value added in the generated documentation that isn’t directly available from documentation comments, such as type members and protocol conformances (though a LLM could certainly suss that out with the right context).
This is awesome and timely for me...going to give it a whirl. Thanks for building. Also, there should totally be an easter egg where clicking something somewhere plays the sound!
This is really cool, but also totally something you'd think existing AI agents should have zero issues with. _Especially_ if they're supposed to be for coding, I'd expect loads of documentation to be baked-in, so to speak
I think it's safe to assume most big players have browser rendering enabled (I hope so). imo AI is struggling with a lot of languages that are not as popular as javascript, mostly because it's more niche and you don't get a lot of good examples on the web.
In my experience, coding agents seem to do a normal fetch when provided links. Which makes sense — headless browser automation is expensive, and only really necessary for interacting with a webpage.
But with these RLHF'd AIs, being confident and helpful as they are, it took me a while to realize that they couldn't actually read the Apple developer links I was giving them. Like a kid who can't read the chalkboard, but doesn't realize they need glasses.
It's expensive for sure, but probably a drop in the water compared to the cost of renting H100. Plus, it would be a massive boost in terms of data quality/quantity for them. But maybe you are right, I'm just surprised it's not the case
Hi Saagar! One could indeed use this method to export this information en-masse, however that would require the kind of automated access that the site ToS explicitly forbids.
For the Swift standard library and other open-source frameworks, you could probably extract this from documentation comments using DocC, Jazzy, or the like.
Once upon a time, I built a project called `swift-doc`, which eventually got Sherlocked. I think what I was most upset about was their decision to call their thing "DocC". Like, adding redundant consonants to avoid name collisions is my shtick.
It is for the same reasons LLM are struggling to produce something that compile in Rust? I was under the impression that most of Rust documentation was plain HTML.
Most of the docs about the language itself live on swift.org (e.g. https://docs.swift.org/swift-book/documentation/the-swift-pr...). I've had pretty good luck getting Claude to write modern Swift code by saying things like "Use Swift 6 structured concurrency instead of GCD". But I could totally see expanding sosumi to include swift.org content, too.
The AI only sees a bit of HTML plus a bunch of JS that, when executed, generates more HTML. If the AI does not run the JS it won’t see everything. During training they probably use a crawler that runs a headless browser behind the scenes to get everything a human would get.
So... The answer is to use during the real-time access the same headless browser as they used during the training? Which they already do, unless you ask specifically to write and run a python script that uses simple requests?
It is like generating static webpages just for SEO: obsolete since 2012[1], and few years later for other major websites.
I think it's because, in the world of programming, people are incredibly critical. Just try putting absolutely anything out there, and you'll get certain types of people picking it apart.
I remember an early experience in my working career, when someone was sharing their sample code with a group, to demonstrate a particular concept. And one of those present picked them up on their use of magic numbers, as if that was at all relevant in the context.
I don't blame anyone for being wary of showing their work in progress. Painters often don't like their subjects trying to take a sneek peak at their work in progress, as another example.
Just saying, sites like these are also pretty great for accessibility, screen reader users in particular.
I think this one would be slightly better if it rendered that Markdown as simple HTML if accessed through a real browser, but I can imagine even this version being pretty useful.
I think it could also make the "Small web" crowd pretty happy too.
Amen to that. It's funny how virtues like accessibility and good API design matter just as much as ever in this new era. Good ideas don't go out of style.
Could you share what is your motivation behind doing whole domain and web hosting ?
Personally I feel that this whole AI induced problem should even exist in the first place, but even then it is ridiculous, that you have to query some web api to solve this problem, why not just publish parsed and converted to .md set of local files and be done with it.
I agree, it'd be great if Apple provided accessible documentation in the first place. Time was, Apple published self-contained docsets that you could download and read offline.
Apple's ToS pretty explicitly forbid the kind of automation required to download everything. But even if someone did that, it'd only be a snapshot in time. And a lot can can change between OS releases.
As for the hosted web app, I wanted to provide this as a public service. I plan to open source it, so anyone can self-host instead, if they're inclined.
Yeah, pre Swift documentation was very sufficient, dense and locally explorable, since then documentation started to resemble .h copy pasta without comments.
For those wondering about the name, it has a fun story behind it: https://en.wikipedia.org/wiki/Sosumi
Jim Reekes talking about this in a deleted scene from the documentary Welcome to Macintosh: https://www.youtube.com/watch?v=k4J8AF25PjA
There's also a episode on https://www.macintosh.fm about the topic that I can recommend.
When I was a kid I used Sosumi as my chime on the apple IIGS!
"AI-readable Markdown" — can't we just say "Markdown"? I'm serious about this, why are we focussing on making things accessible to AI when they should just be machine-readable and accessible to human beings in the first place? No need to taint this by bringing AI into it.
> why are we focussing on making things accessible to AI
Because that’s the authors actual goal? To take a web page that looks fine to human eyes but is unintuitively not accessible to AI. That’s genuinely useful and valuable.
Sure it’s no different than converting it to markdown for human eyes. But it’s important to be clear about not just WHAT but also WHY.
C’mon now. This isn’t controversial or even bad.
Makes one wonder what apple’s actual goal is
I mean.... it could have broader appeal without artificially restricting its audience
Not everything is for everyone. Y'all are criticising this project's bean soup because you don't like beans.
If you think it has less appeal because it was built for and advertised for AI that’s a you problem.
Great promise; sometimes need to reference docs to build context.
I looked at the examples you posted and did a quick glance. For example
'''init?(exactly: Float80)'''
the tool converted it to
'''- [initexactly-63925](/documentation/Swift/Double/init(exactly:)-63925)'''
To achieve its goal I would be worried that it dropped the verbatim function signature. Claude still figured it out, but for more obscure stuff that could be an issue.
Thanks for pointing that out. That’s most likely a mistake in how I’m translating into Markdown. I’ll look into this.
Following up — I just pushed a fix for this. This latest version significantly improves how references like protocol conformances and default implementations are rendered.
Nice! Do you think it could be adapted to other docs sites?
I made a small clone of the tutorials section (https://clone-swiftui-tutorial.vercel.app/) where the content is already Markdown (and use codehike to turn the markdown into a rich UI). This made me realize that codehike is AI-friendly, in the sense that even for non-linear UIs the original content is still AI-readable Markdown.
How to reliably HTML to MD for any page on the internet? I remember struggling with this in the past
How hard would it be to build an MCP that's basically a proxy for web search except it always tries to build the markdown version of the web pages instead of passing HTML?
Basically Sosumi.ai but instead of working on only for Apple docs it works for any web page (including every doc on the internet)
I think HTML -> Markdown is a bit of a red herring.
In many cases, a Markdown distillation of HTML can improve the signal-to-noise ratio — especially for sites mired in <div> tag soup (intentionally or not). But that's an optimization for token efficiency; LLMs can usually figure things out.
The motivation behind Sosumi is better understood as a matter of accessibility. The way AI assistants typically fetch content from the web precluded them from getting any useful information from developer.apple.com.
You could start to solve the generalized problem with an MCP that 1) used a headless browser to access content for sites that require JS, and 2) used sampling (i.e. having a tool use the host LLM) to summarize / distill HTML.
https://pure.md is exactly what you're looking for.
But stripping complex formats like html & pdf down to simple markdown is a hard problem. It's nearly impossible to infer what the rendered page looks like by looking at the raw html / pdf code. https://github.com/mozilla/readability helps but it often breaks down over unconventional div structures. I heard the state of the art solution is using multimodal LLM OCR to really look at the rendered page and rewrite the thing in markdown.
Which makes me wonder: how did OpenAI make their model read pdf, docx and images at all?
There are APIs such as Jina AI's reader API that do this pretty well. It doesn't produce output as clean as Sosumi for Apple docs, but it's free and does a decent job.
https://jina.ai/reader
https://github.com/vivekVells/mcp-pandoc
Hm, I would have extracted the markdown from the Swift source code. That’s what Apple uses to generate their pages, using https://www.swift.org/documentation/docc/.
For example, AFAIK, https://github.com/swiftlang/swift/blob/main/stdlib/public/c... is used to generate https://developer.apple.com/documentation/swift/array.
This is only possible for some of Apple‘s open source Swift code, including the Swift standard library. This is not the case for hundreds of other SDK frameworks, such as SwiftUI.
Even for those open source projects, there is still some value added in the generated documentation that isn’t directly available from documentation comments, such as type members and protocol conformances (though a LLM could certainly suss that out with the right context).
This is awesome and timely for me...going to give it a whirl. Thanks for building. Also, there should totally be an easter egg where clicking something somewhere plays the sound!
Great idea! I just added that. Try clicking the icon in the header.
Ha, amazing and so nostalgic
As someone who is currently building my first iOS app, I am extremely happy to have this. This will be much nicer doing my animation documentation.
This is really cool, but also totally something you'd think existing AI agents should have zero issues with. _Especially_ if they're supposed to be for coding, I'd expect loads of documentation to be baked-in, so to speak
Hope this make apple's Xcode team realize they should do this, especially with all the recent AI integration.
I think it's safe to assume most big players have browser rendering enabled (I hope so). imo AI is struggling with a lot of languages that are not as popular as javascript, mostly because it's more niche and you don't get a lot of good examples on the web.
In my experience, coding agents seem to do a normal fetch when provided links. Which makes sense — headless browser automation is expensive, and only really necessary for interacting with a webpage.
But with these RLHF'd AIs, being confident and helpful as they are, it took me a while to realize that they couldn't actually read the Apple developer links I was giving them. Like a kid who can't read the chalkboard, but doesn't realize they need glasses.
It's expensive for sure, but probably a drop in the water compared to the cost of renting H100. Plus, it would be a massive boost in terms of data quality/quantity for them. But maybe you are right, I'm just surprised it's not the case
Reminds me a bit of https://www.felixcloutier.com/x86/ - turning thousands of pages of Intel PDF into more semantic documentation.
Is it possible to download an archive of the data so I can run searches against it locally (without AI)?
Hi Saagar! One could indeed use this method to export this information en-masse, however that would require the kind of automated access that the site ToS explicitly forbids.
For the Swift standard library and other open-source frameworks, you could probably extract this from documentation comments using DocC, Jazzy, or the like.
An “AI” that makes text “AI”-readable. How does that make any sense?
I don't even bother trying to render docc catalogs into JS. It's a royal pain that breaks easily.
If GitHub could support .docc files, that would be great. Otherwise, I still use Jazzy Docs.
Once upon a time, I built a project called `swift-doc`, which eventually got Sherlocked. I think what I was most upset about was their decision to call their thing "DocC". Like, adding redundant consonants to avoid name collisions is my shtick.
Long live Jazzy.
I love your stuff!
Thanks!
It is for the same reasons LLM are struggling to produce something that compile in Rust? I was under the impression that most of Rust documentation was plain HTML.
Very interesting. You have any before and after examples?
Curious how it handles some of the concurrency stuff. Actors, async/await etc..
Sure, here are a few more examples:
- https://developer.apple.com/documentation/swift/tasklocal
- https://sosumi.ai/documentation/swiftui/button
- https://sosumi.ai/documentation/foundation/measurement
Most of the docs about the language itself live on swift.org (e.g. https://docs.swift.org/swift-book/documentation/the-swift-pr...). I've had pretty good luck getting Claude to write modern Swift code by saying things like "Use Swift 6 structured concurrency instead of GCD". But I could totally see expanding sosumi to include swift.org content, too.
Edit: Another Swift developer life hack: for new language features, copy-pasting the Swift Evolution proposal works pretty well! https://github.com/swiftlang/swift-evolution/tree/main/propo...
Are you making use of the new markdown files Apple started shipping inside Xcode? They’re probably copyrighted though
I’m building a swift app now and will most definitely give this a try
Just wondering - can't AI read HTML? If so how are we training our models?
The AI only sees a bit of HTML plus a bunch of JS that, when executed, generates more HTML. If the AI does not run the JS it won’t see everything. During training they probably use a crawler that runs a headless browser behind the scenes to get everything a human would get.
So... The answer is to use during the real-time access the same headless browser as they used during the training? Which they already do, unless you ask specifically to write and run a python script that uses simple requests?
It is like generating static webpages just for SEO: obsolete since 2012[1], and few years later for other major websites.
[1] https://www.i-programmer.info/news/81-web-general/4248-googl...
Ai can’t read something dynamically rendered with JavaScript. At the moment.
They can, but the token to content ratio is far less, so they work less effectively when it's put into the inference context window.
Another awesome project that does this for Apple's docs: https://llm.codes/
Also, Apple has started shipping docs like this, too. They are a bit hidden but you can find them here:
/Applications/Xcode-beta.app/Contents/PlugIns/IDEIntelligenceChat.framework/Versions/A/Resources/AdditionalDocumentation
Do you have a public repo? Would love to see how it's working.
Not yet, but I plan to open source it soon. Just gotta tidy up a little bit, you know?
Why do people always say this?
It’s ok to just start coding with a public repo. Code isn’t a secret.
I think it's because, in the world of programming, people are incredibly critical. Just try putting absolutely anything out there, and you'll get certain types of people picking it apart.
I remember an early experience in my working career, when someone was sharing their sample code with a group, to demonstrate a particular concept. And one of those present picked them up on their use of magic numbers, as if that was at all relevant in the context.
I don't blame anyone for being wary of showing their work in progress. Painters often don't like their subjects trying to take a sneek peak at their work in progress, as another example.
- Keys/tokens in the code. Yes, it’s bad practice but for a hobby/personal project it’s not the end of the world
- Not wanting to get roasted
- Open source = dealing with a lot of entitlement
And the list goes on. Putting code out into the world (publicly) often sets you up for future obligation of some kind (even if it’s just saying “no”).
None of this is a stance against open source, but I understand where people are coming from.
Can't wait to use this. Its awesome!
Is it open source?
This is awesome; great work.
Wasn’t one of the benefits of AI that we don’t need special documents for AI to understand the data?
Understand yes but the issue that this project is built for is to access data
sosumi.md seems to be a better fit domain-wise, no?
Love the name.
Just saying, sites like these are also pretty great for accessibility, screen reader users in particular.
I think this one would be slightly better if it rendered that Markdown as simple HTML if accessed through a real browser, but I can imagine even this version being pretty useful.
I think it could also make the "Small web" crowd pretty happy too.
Amen to that. It's funny how virtues like accessibility and good API design matter just as much as ever in this new era. Good ideas don't go out of style.
Aha, another developer doing Apple's job.
> "Ever notice Claude struggling to write Swift code?"
Yes, that is why I quit using Claude and swapped to ChatGPT about a year ago. I've had substantially less issues with GPT.
I remember copilot with GPT helping me export Swift to C for interop. Something I could hardly find good examples and info on in my google searches.
ChatGPT does swift and objc remarkably well including UIKit and core data. Its gotten much better for project files with gpt5-mini
Could you share what is your motivation behind doing whole domain and web hosting ?
Personally I feel that this whole AI induced problem should even exist in the first place, but even then it is ridiculous, that you have to query some web api to solve this problem, why not just publish parsed and converted to .md set of local files and be done with it.
I agree, it'd be great if Apple provided accessible documentation in the first place. Time was, Apple published self-contained docsets that you could download and read offline.
Apple's ToS pretty explicitly forbid the kind of automation required to download everything. But even if someone did that, it'd only be a snapshot in time. And a lot can can change between OS releases.
As for the hosted web app, I wanted to provide this as a public service. I plan to open source it, so anyone can self-host instead, if they're inclined.
Yeah, pre Swift documentation was very sufficient, dense and locally explorable, since then documentation started to resemble .h copy pasta without comments.