Traefik is pretty cool, but suffers from the same, terrible problem of Ansible: there is a lot of documentation, and a lot of words written, yet you can never find anything you need.
I have used it since v1 and I routinely get lost in their docs, and get immensely frustrated. I have been using Caddy for smaller projects simply because its documentation is not as terrible (though not great by any stretch)
Technical writers: documentation by example is good only for newbies skimming through. People familiar with your product need a reference and exhaustive lists, not explanation for different fields spread over 10 tutorial pages. Focus on those that use the product day in and day out, not solely on the "onboarding" procedure.
This is my pet peeve and the reason why I hate using Ansible so damn much, and Traefik to a lesser extent.
One of the problems that the yaml interpreter class of languages, or whatever you'd call them, suffer from is the fact that yaml itself is a language and tends to be more or less undocumented in the interpreter docs.
It's sort of assumed that you are going to do extremely simple tasks on very flat data structures. That doesn't tend to be the reality that most of us live in. And to really get the most out of these languages you have to understand an entire unspoken set of rules on how to use yaml. That's never really pointed out in the docs.
Additionally, there are docs for the unique settings for each module but as far as using the standard settings, additionally, its rarely clear how to operate on the data that might be returned or combined with anything mildly complex, you are given a dozen 1 stanza examples for each item like a stack of ingredients and then told to bake a cake.
I've had this experience with basically every one of the various yaml interpreter systems I've used.
After a few 100k lines of yaml I can get things done but the docs are useless other than a listing of settings.
Isn’t it why toml is seemingly increasingly used to replace yaml in projects?
I hope not, toml is even worse at complex things and just slightly better at the stuff that isn't confusing. Add a k:v to a mildly complex dict.
At this point, I'm pushing into a place where I'm just going to switch to go because its getting to be a mess.
It’s insanely better at config.
It’s about as bad at being a programming language or data structure serialization format, though.
But then, what is a config file if not a representation of a data structure?
With the restriction that it has to be represented in human-friendly text.
JSON’s a crazy-bad serialization format, too, for that matter. It doesn’t even know what a damn integer is.
But yaml is fine at config, it sucks at looping, conditionals and data structures, if you aren't fixing that its just another standard we have to learn, so thanks for that
No, it's not fine at config. It's a bag of weird corner cases and incompatible revisions which you have to debug instead of doing your actual work.
Here's a good summary: https://ruudvanasseldonk.com/2023/01/11/the-yaml-document-fr...
Nope it isn't.
There are so many things that aren't expressible in TOML that any anywhat complex system will want... it's not even a contender.
So, one problem a lot of configurations are trying to solve: modularity. I.e. how to allow different actors to change the parts of the configuration they want. Everything under /etc nowadays is of a form /etc/*.d/*.* that is all configurations are directories of multiple files with some ridiculous rules (like "prefix file names with digits so that they sort and apply in the "right" order etc.) XML had a better approach with namespaces and schema, but maybe not perfect.
Polymorphism. Any non-trivial configuration system will have plenty of repeating parts. NetworkManager connection configurations? -- They are all derived from the same "template". Systemd device services -- same thing, they are all coming from the same "template". There are plenty more examples of this. But, languages like YAML or TOML don't have a concept of polymorphism in them. This is never encoded in the configuration itself. Instead, every tool that needs to be configured and needs some degree of polymorphism rolls its own version.
Constraints. It's often impossible to describe the desired configuration through specifying the exact values. Often the goal can be described as "less than" or "given the value of X, Y should be multiple of X" and so on. Such concepts, are, again, not expressible in TOML or YAML and friends.
NB. Types are a kind of constraints.
Identity. It's often necessary in configuration to distinguish between two sub-sections that look the same and two sub-sections that designate the same exact object. Like, when configuring VMs with eg. disks: are they supposed to mount the same disk, or does each VM need a separate disk, that just has the same physical characteristics?
In my experience toml is worse at anything complex. It's nice as an .ini replacement but makes even yaml look sane in comparison if you want to use it for very complex or deeply nested stuff. But it wasn't designed to do that anyways
Am I alone in greatly preferring nesting in toml compared with yaml?
Did you try to convert any mid-complexity ansible role into toml? It was very interesting exercise for me and vastly conclusive.
Toml is great for simple use-cases. For complex ones you have the same problem that yaml has: Templating a language with significant whitespace via text substitution is a horrible horrible idea. Somehow this sad state of affairs has become industry standard.
Its not even the white space, a food liter or language server can handle that. It's not that, as much ad the fact that the most complex data structure is a list.
If you want to get crazy, you can push a dict into a list and operated on it but it gets tough at the second level. And don't get me started on if/else statements.
To illustrate this point, here is how to have a multi line value in yaml: just kidding, it’s so confusing that there is a whole website to help you figure it out: https://yaml-multiline.info/
Those examples don’t look confusing except there being more than one way.
Good luck remembering which is which :)
I honestly wonder why not just write your web server in node or something. It would be traceable and testable and probably performant enough. There's just so much arcana inside platforms like traefik or nginx where they do all this miraculous stuff if you just add the right flags, but also when it doesn't work it's a total black box and there's no way to discover what it thinks it's doing.
This take is, at best, disingenuous, and at worst, dangerous. The Traefik maintainers and community contributors (including myself) have collectively invested hundreds of man-hours writing and improving documentation, specifically in response to feedback from users that things are hard, unintuitive, or complex.
You are discounting massive amounts of unpaid labor done specifically for people like you. At this point, if you can't find what you're looking for, it's on you. Maybe do a little bit of your own homework instead of throwing your hands up after 2 minutes and crying to the maintainers.
Investing a lot of time and trying really hard is not the same as adding a lot of value. If your users don't find value in your documentation, saying "But we spent a lot of time on it!" doesn't really change anything.
And, to be clear, I have no idea if the person you're responding to's criticism is valid. But I also know that your response does not negate their criticism at all.
How about submitting a PR to improve the documentation instead of complaining about it?
Customer: I can't find anything I'm looking for in your store.
Store: I spent a lot of time arranging things around in the store, if you can't find what you're looking for you can stop complaining and write signs for us.
Customer: Or I can just use your competitor who actually cares about their customers. ¯\_(ツ)_/¯
Except the relationship between stores, customers and competitors all revolves around money and doesn't make sense in this context.
A PR needs maintainer approval. And the maintainer I've seen thinks the documents are good enough already. In cases like that, a PR might not be able to solve the problem.
Complaining about it reroutes people to better projects, and pushes the project to fix the problem.
Your adversarialism isn't a good look. Users are allowed to have opinions, this does not mean they have maintain the work themselves.
This is why OSS looks like a cult at times. People are allowed to criticize your project and complain about it. They have no obligation to become a contributor. "Submit a PR" is such a conversation killer.
I didn't use try Traefik's documentation, but the complains appear to be somewhat structural. Meaning a PR would need to possibly restructure at least part of the documentation, or add a whole section of documentation of a different type.
You can't expect someone not core to a project to just propose to restructure the whole documentation. Which may also mean changing the website.
And in any case, such overhaul coming from a "nobody" would very likely be rejected as being both too large or incomplete or not desirable.
Re-structuring needs to be pushed for by at least one person from the core team.
So yeah "Just submit a PR" in that context is not an answer, it's an excuse to avoid trying to understand the problem and actually improve the situation.
I never used Traefik and have no opinion on it one way or the other as such. But if this is the response to some criticism of the documentation – which you can agree or disagree with, then you've done more to turn me of from Traefik than anything anyone here can write.
Note that the commenter isn't associated with traefik itself
Right so; the way it was written implied that at least they had spent significant amount of time on it, at least to me (I tried a quick grep in "git log" to see if I could find them, but couldn't find anyone matching the username here).
Disagree that this isnt a generic problem. And i'll take the same amount of umbrage at you calling it disingenuous. There are dual needs here. Having to read a story and take in a wholly unrelated workflow just to discover only half of the switches available to the feature im looking up is a problem.
And when there isn't just straight documenting of what's been implemented then it is an unreasonable gate to usage which limits customers to only the flows imagined by the technical writer.
Which itself breeds this sort of refusal to participate. Either the end user is ungrateful and needs to express that gratitude through silence or there's a smug moderator who's read everything and knows which paragraph of which tutorial has the answer and harangues anyone asking with a link and a "why didnt you read sentence 5 of paragraph 2 of a tutorial written 2 years and 3 major versions ago?"
Yes, those are two different kinds of documents for two different audiences.
You need both a java doc style big 'ol list of every function and what they do, and a narrative/workflow/primer based section of you documention
My latest gripe in this category - opentelemetry. Thousands of pages. Very little about actually achieving basic common workflows.
Oh boy that hits home. Been deep in the OTEL world the last months and the official docs are very, very undercooked.
The best OTel docs I've seen have been from observability vendors. The CNCF needs some volunteer technical writers or something. I think a lot of those docs suffer from being written by people who know the spec inside and out.
Same experience. Otel is one of the wordiest docs I've ever come across that says very little.
Further, I found a lot of little bugs that are hard to Google, or when Googling finding open issues that are either known and working on, or no response at all.
I ended up just throwing it in the garbage and using direct connectors. I like what Otel is trying to achieve, but it feels extremely opaque and half baked at the moment.
Yes! Otel, is exactly what I was thinking. It’s for people working on it with access to tribal knowledge. For mortals working to add it to their service it’s not fit for purpose
Oh man, I FEEL this comment. That was one absurdly awful set of documentation, because they not just have a lot of confusingly placed repeat content, they also follow the philosophy of only explaining top level initial conceptual primer for everything, and only explaining the actual main use case of the component 3 navigation pages deep.
So a beginner has to jump a BUNCH of pages to get a primer, and an expert has to bookmark the couple actually-useful pages and later give up and just look at github for specific operators/processors when they already know the basic config inside out.
Several things I needed were in completely separate documents not linked together.
When I started using Honeycomb, I had such a wonderful integration experience with their Beeline SDKs.
Then they transitioned to OpenTelemetry – for very good, justifiable, "good community member" reasons – and yikes, everything got so much more complicated. We ended up writing our own moral equivalent to the Beeline SDK. (And Honeycomb have followed up since with their own wrappers.)
There's so much I love about Open Source, but piles and piles of wildly generic, unopinionated code... ooft. :-)
I feel like it is endemic to anything OPS / DevOPS. Lot of uselessly verbose "documentation" but no list of whatever you really need.
All in the name of selling products which abstract those parts, consulting or courses.
Do not agree WRT ansible, been using it for well over 5 years and usually a google search points me right at the correct part of the documentation to answer my question. Ansible, the tool itself, can be a bit obtuse, largely IMHO because of the YAML source language, so some concepts are hard to translate into the tool, but the documentation has never bothered me.
As far as "a lot of words written, can't find what you need", Fortinet is my poster child there (based on trying to use it a decade ago). Everything I looked up there had 10,20,30 pages of introductory material with the Fortinet stuff spread throughout it.
Alright, please link me to an exhaustive list of Jinja filters supported by Ansible out of the box. I'll wait.
What you are given is https://docs.ansible.com/ansible/latest/playbook_guide/playb... and you need basically to read/scan each example until you find what you need [1]. Do you call that good, especially when these are basically the only way of doing anything a little complex? That's a sure way of killing my flow and productivity in its tracks. I have been through this page in anger a dozen times, and I still have no idea what Ansible filters can or cannot do.
Also, using Google to find stuff is "cheating". The goal of documentation is to be able to use it as reference; if you need an external tool to find anything in it, that defeats its purpose a bit. When people wrote documentation books, they had to make sure it's usable, legible and efficient. These days apparently that's become lost art.
1: these examples are not even exhaustive, because they don't list all the builtin Jinja filters; chances are that what you need isn't listed on that page, but you should instead refer to https://tedboy.github.io/jinja2/templ14.html
I like that it forces users to read the docs to find the functionality. Users don't read the docs, and then they wander around the internet looking for a random blog post with a snippet for one problem, and they don't ever really learn how to use the program.
Users are a bit like high school students just skimming books for an answer to fill in on a test. They need to be forced to learn.
This doesn't make a lot of sense in the context of the parents. Did you post this to the wrong parent? To accomplish what you are asking, a project needs actual good documentation. Everyone has agreed that is good. The only real disagreement here is whether Ansible docs have this, and regardless whether they do, they definitely have the example-driven docs that I think you are saying you don't think should exist, so you definitely aren't supporting the Ansible status quo.
The rest makes more sense if you assume parent post meant to write "read the code".
As you say, Ansible's filter list does not include the base Jinja2 filters, which I guess is a difference of opinion. I feel that is preferred to reproducing the Jinja2 documentation, especially as the Jinja2 filter list is the first (non-TOC) link on the page.
Also going to disagree about "using Google is cheating". The purpose of documentation is to help me get stuff done. The Internet is not printed on dead trees, I don't want to read through a TOC or index looking for what I want when I'm searching, I want to use a search engine. I often don't want a reference, I want to quickly find how to do something. I rarely want to read about all the filters, instead I want to find the even/odd filter, or the the default or omit filter. Yes, sometimes I want to brush up on all available filters, but that's rare.
I'm not GP, but I agree with both them and you so thought I'd chime in.
You're absolutely right that there are big omissions/holes in the Ansible docs, but I also think that using Google is not "cheating." My ideal of great documentation sounds like exactly what you would agree with: A complete and comprehensive "book" (could become a physical printed book, but needn't have to as it should be equally usable with good old-fashioned hyperlinks). It should have a logical flow, introductory sections to describe pre-requisite knowledge/concepts and things that are broadly applicable to the project as a whole. It should have a table of contents, and it should definitely have an index and comprehensive lists/tables of API details such as available field/properties, which options are valid (for enum fields), etc. Your example of Jinja filters supported by Ansible is a great one. I really miss the 90's era here where such manuals were common practice, even for things like PCs.
With that ideal described, though, I think it's important to recognize pragmatism and feasibility. Documentation takes time and money to produce. Search tools (including Google) already exist and can provide a valuable addition without spending time/effort on it, so I think they should be used. That said, I agree that it's not a good idea for doc writers to rely on that for things to be found! Table of contents, logical flow, and indexes should absolutely be thought through. If the documentation is just a bunch of random unorganized and uncatalogued pages that can only be found with a search engine, that is really bad and they should feel bad.
I think Ansible falls right in the middle there. It undoubtedly has some real glaring omissions/holes in it, but it is also not nearly the worst I've seen as well. I do dread having to go the Ansible docs though, which is an indictment against their quality, and the more I think/write about this the more I agree with you lol.
Maybe you use ansible in a more complex environment than I do. I use ansible to provision servers, and jinja2 filters are some of the least-used things for me. I try to keep my ansible roles short and simple, and needing obscure jinja2 filters is a clue to me that I might be getting too fancy.
That said, when I have a question about jinja2, I find that the jinja2 docs are better than the ansible docs.
I agree. We all would benefit by giving more exposure to documentation frameworks such as https://diataxis.fr
Also https://docs.divio.com/documentation-system/
These are "the same". Diátaxis is the fork the author made after (I presume) they left Divio.
MSDN also follows these principles
Discussed on HN: https://news.ycombinator.com/item?id=33721314
Thank you for the link <3
I'm glad to have clicked through for curiosity's sake. Diátaxis is tremendously interesting.
For folks who might recognize the author's name:
Daniele Procida: Director of Engineering at Canonical. Creator of Diátaxis and BrachioGraph. Django core developer. Fellow of the Python Software Foundation.
I usually feed all the documentation to an llm and ask the question to the llm that fetches the info with rag. For example:
Configuring Traefik to proxy traffic to multiple backend services is a common use case. You can achieve this by defining multiple entry points and routers in Traefik's configuration, each pointing to a different backend service. Here's a basic example of how you can configure Traefik to proxy traffic to two backend services:
[...more stuff...]This might be an additional way to quickly iterate on the content of the documentation itself. If I were the one writing, an easy test is passing the documentation to a lay person and asking them if they have what they need to perform X by following the documentation.
Perhaps having a focused LLM generate the steps could help catch some documentation deficiencies.
What kind of documentation is this though? Is this how to bake a cake or tie a necktie, or is it how to setup a reverse proxy for the services in your k8s cluster?
If it's something a lay-person could do then I think this is a good strategy (though depending on the size/scope of the project/documentaiton it does seem like a pretty big effort to undertake without compensation), but if it's something highly technical like Traefik, I expect a lay-person to not even understand half the words/vocabulary in the documentation, let alone be able to perform X by reading it and following it.
If we push this into the software development domain, my expectation would be something like "documentation should allow a software developer to do this thing without knowing the underlying tooling".
Going off this specifically. I don't know how to do this. I actually have a k8s cluster on a home server waiting for me to do exactly this. Ideally there would be a doc that would start with ingress selection, and then guide a user through how to get it set up with common use-cases. Or something like that. Like others in this conversation, I've been leveraging LLMs with varying degrees of success to try and navigate this.
Can you describe more on your process? Which LLM are you using? Are you doing soething specific to make it us RAG or is that automagic (might be obvious depending on which LLM you are using but)? How do you feed the documentation in? for example, when the documentation has more than one page, how do you get that content in to the LLM? Is it part of the prompt or something you've tuned it on? have to clone the docs site, turn it into plan text and feed that in to the prompt or can you pass a URL and have it crawl ahead of time or something?
This is the system I've been dreaming about but haven't had time to dig into yet. I've got ollama and openwebui set up now though, and with OpenAI getting bigger context windows it seems like it might be possible to inject the whole set of docs into the prompt, but I'm not sure how to go about that
I did this with Tailscale, which I have endless problems getting to work reliably. Their documentation is a joke. The process is pretty simple: scrape, than concatenate into a large text file and submit.
Pydantic falls into this box for me. The maintainer refuses to build API reference documentation, as they feel that there should only be one source of information. It's their project, of course, but every time I need to find a method on an object, I am scouring pages of prose for it. Sometimes it's just easier to read the source.
What's missing from the existing API documentation?
https://docs.pydantic.dev/latest/api/base_model/
Oh nice! I'll admit, my comment was based on outdated knowledge (the last time I used Pyandtic for a major project was in 1.x).
https://github.com/pydantic/pydantic/issues/1339
Haproxy does the whole documentation side of things very well.
The docs are very straightforward and thorough.
same reason why Terraform AWS Provider is better documentation than AWS documentation
https://registry.terraform.io/providers/hashicorp/aws/latest...
If I can't find the answer to what I need there I usually resort to LLMs, they are surprisingly good and fetching the info you need out of these massive documentations. The failure rate is quite high though so a lot of trial and error required, but the LLM at least gives you some hints to where to look for it.
My primary use case for LLMs so far has indeed been to avoid terrible technical documentation.
I'm on the lookout for an LLM that is relatively small and focused on technical documentation. I don't need it to write prose or haiku, just answer my questions based on documentation
Yep. Build it into an editor and I'd be the first customer.
Well said, this extends far beyond Traefik. Far too much documentation these days is tailored for people who have never used software of it's type. This was a workable strategy during the Great Developer Boom, but that's more or less over now.
As a developer who didn't come from this Boom, I have been constantly frustrated by this trope, and I hope the changes in the industry will tip the scales back toward solid reference documentation, so that I can feel confident in deploying more of these technologies.
Putting that more general note aside, I have been a Traefik user for years and I do recommend it. But a lot of what it does is difficult to cite using solid docs.
I've often resorted to just looking through the project on github and finding whatever source file is responsible for parsing the configuration files to figure out what each option does.
This applies nearly universally in open source. If documentation isn’t sufficient, jump into the source code and most things can be figured out. Reading other people’s code is also a great way to learn/experience different approaches
Ansible is definitely requiring constant lookup in the documentation.
I've found a pretty good workflow with using ansible-doc though, with two-three aliases that I use constantly:
Then I'll:1. Use adls to quickly search (in less with vim bindings) for relevant commands,
2. Check up the docs with `ad <command>`.
3. Almost always immediately jump to the end (`G`) to see the examples, which typically provides a direct answer to what I need.
Since authoring Ansible scripts is so dependent on the docs, I think they really should make this process work better out of the box though, providing some interface to do these lookups quicker without a lot of custom aliases.
I keep the module index[0] in my bookmarks bar and that's also been pretty easy to search and read.
[0] https://docs.ansible.com/ansible/latest/collections/index_mo...
Possibly unpopular opinions to follow. This is made even worse by:
- Having documentation split between v1 and v2, that is similar yet different enough to yield half-baked configurations before you realize what you did wrong. The website itself provides the barest of subtle changes to distinguish the two. Edit: I learned all this prior to v3.
- Supporting multiple config formats (TOML and YAML) which makes it that much harder to hunt down relevant forum posts for support. That wouldn't be a huge problem if it weren't for things that you need that aren't in the documentation (above)
- Multiple configuration _modes_. You can set options on the CLI, or in config files, and they are not 100% reflected between the two; some things must be in config files no matter what. Config files themselves are split between "dynamic" and "static" configs, and you must place the correct options in the right file.
- The one thing that Traefik does well is routing traffic to/from containers. Container labels are used to drive this. How to map those label namespaces to various internal and user provided aspects of Traefik is unclear in the docs, at best.
- Traefik exposes a dashboard web GUI by default. Yet much of the useful diagnostic and troubleshooting information is only found in the container logs.
Retiring v1 completely, picking a single configuration language/mode, and providing a rich library of example docker-compose configs, would go a very long way to securing this project's future.
The documentation split is unfortunate and the GUI really just a status page. The other points are a strength. A pattern that works well: Put all Traefik config in your Docker container definitions, as command line flags and labels, plus the dynamic config provided as a volume. That gives you all the flexibility and only one or two places to look for the config (e.g. a Docker compose file and the dynamic config file)
I share the same sentiments. I dread having to go through Ansible docs because it's so densely packed. Meanwhile Caddy's docs feel too sparse, and too many spread out tutorials. The reference isn't well thought out either imo.
We're revamping the Caddy docs this year.
Funny you say that because we don’t have nearly any examples in the Caddy docs. We’re working on improving them later this year.
Examples are good in docs. But documentation that's only made of examples and tutorials... not so much.
Thanks for Caddy btw. Neat little tool.
It's funny because one thing I like about ansible is how easy it is to get the reference doc for any module with `ansible-doc -t module`.
I do sometimes struggle to find the right doc when I'm searching for something about ansible core itself, but that doesn't happen too often.
Oh man are you on to something!!! One huge, bad side effect of web is the atomization of an overall body of work into 62.9 million links.
One pdf please. The book concept works!
You know who's docs blow too? Mellanox. I hate their stuff.
And to give credit where due: intel does a damn good job.
This is strange. I also don't like the docs but for a different reason.
I would rather have a more examples. And kinda _advanced_ and complex, rather than trivial we see in the docs.
Even though I had a working V1 configs and had a know-how about lingo / architecture like routes / services I still struggled for a day or two to properly configure a pretty simple workflows in v2 like:
* add TLS with LetsEncrypt
* configure multiple domains
* configure multiple services
* add Basic Auth for some domains
That said, more detailed and extensive docs would be much better.
I also remember finding things in github issue comments that worked as bugfix/workaround of something from the docs.
PS. For now I've moved to Caddy for simplicity and better Caddy DSL compared to yaml/label verbose config.
You really need at least three documentation targets:
- onboarding the newbies workflows/tutorials - intermediate "focus on the important bits" workflows/tutorials - exhaustive references
There might be other useful ones as well, but I never see those three hit at the same time adequately.
Those are great points. Even the page layout of the documentation is terrible. Whywe have huge monitors and millions of pixels if I have to read content from a very narrow column, which is a mile long?
Eg: https://doc.traefik.io/traefik/routing/services/
If you visit that page in your desktop browser you'll get less words per column then seeing this on the iPad (works even in dev tools window). Mind blowing.
I'm using the Traefik and have the same experience with the documentation. It can be time consuming to configure it properly if you are not a power user.
I'm happy with it though, it's a great piece of software. I wonder is there any other product out there with a similar feature set?
Some projects need documentation, some need cookbooks. Sounds like traefik is the latter.
Hopefully as an aside (I know very little about traefik so maybe I am talking about them too and don’t know it), it seems like in the time since I abandoned Java they have weaponized that architectural strategy and I have no patience for it. I look at that sort of documentation and my eyes glaze over. Or if they don’t I feel disgust or anger and all three result in my stomping off.
Opentelemetry, particularly the stats code (vs the span code) triggered a lot of this in me. It has several sets of documentation that say different things. It took me a long time to figure out how to connect the amorphous dots, and then I didn’t entirely agree with their solution anyway.
Just another one for collection: conda. Especially the parts about conda-build, meta.yaml etc. There are only examples w/o any way to tell what's available. And the source code is frustratingly twisted, undocumented and all over the place. Something that makes creating conda packages an extremely frustrating experience, to the point that it's significantly easier to create archives and write the metadata by hand than to rely on conda tooling.
If only there were a program that had crawled bazillions of documents, including all of the traefik documentation, examples, and thousands of code files using it, and if only said program were especially designed to answer natural-language queries about said documents.
This was exactly my experience. It’s incredibly frustrating to search documentation only to be stuck with examples that are related, but don’t fit one’s exact situation, and don’t explain the underlying behavior.
I agree that the documentation could be better, but it isn't that bad. I enjoyed all the gophers, and these images really helped me understand the structure.
However, I find it amusing that you wish there was a better reference. I think getting to the initial setup is quite hard. Once you have that, extending it is straightforward.
I don't think ansible docs are that bad.
I use duckduckgo and adding !ansible to my search usually gets me what I need pretty directly.
From this point of view the Oracle RDBMS handbooks ca. 1998 were pretty good. Come to think of it, they were pretty good all around.
I want both, in the same page if possible, for every possible permutation of input arguments. In theory ansible does this, but then it doesn't link to "you might use it in combination with...", essentially, it lacks integration of multiple things in the reference docs. but I didn't find ansible docs that bad? Most of the time I search module name and find the reference doc