Copilot for Docs

https://githubnext.com/projects/copilot-for-docs/

by irakeshpurohit • 3 years ago

108 53 3 years ago

githubnext.com

joeblubaugh2• 3 years ago

Love how 3/4 reactions are engineers from GitHub rather than customers.

danielmarkbruce3• 3 years ago

They are engineering tools... if your internal engineers are loving your engineering tools, you are onto something.

joeblubaugh2• 3 years ago

Maybe, but the marketing message comes across as “GitHub says GitHub product is good”

switch007• 3 years ago

I'm so used to links not being underlined, I tried clicking the orange 'Copilot for docs' about 4 times !

linkmangotduped• 3 years ago

Me too :'(

jskherman• 3 years ago

username checks out... but was created 3 hours ago

cbovis• 3 years ago

Even the text which is underlined isn't a link lol.

wccrawford• 3 years ago

Underlined, and has an on-hover color change!

chrismorgan2• 3 years ago

I'm so used to links not being underlined

Here’s a user stylesheet I’ve been using for 2½ years (when color-mix() landed behind a pref in Firefox Nightly!):

  :any-link {
      text-decoration: underline color-mix(in srgb, currentcolor 30%, transparent) !important;
  }

  :any-link:is(:hover, :active, :focus) {
      text-decoration: underline !important;
  }

This means links get a semitransparent underline normally, and full-opacity on hover. I reckon it’s an excellent balance. Occasionally you’ll get double underlining due to people using border-bottom instead of text-decoration for some unfathomable reason, and occasionally there’ll be link-styled buttons that won’t get this underline, but all up I’ve found it pretty good as an intervention.

(I’ve been using this technique on websites I make since 2019, though I haven’t yet had the opportunity to use color-mix() on a public site, which only stabilised in browsers 6–11 months ago. My preferred technique there will be `:any-link:where(:not(:hover, :active, :focus)) { text-decoration-color: color-mix(in srgb, currentcolor 30%, transparent); }`.)

pbowyer• 3 years ago

This is a really nice technique, thanks for sharing.

Question: why would you prefer the latter technique on public sites, vs what's in your user stylesheet?

r0ckarong• 3 years ago

Great, all I ever wanted was a hallucinated manual.

number62• 3 years ago

1. Hallucinate the Manual;

2. Hallucinate Tests based on the Manual;

3. Hallucinate the Code to pass the Tests;

Voila! You hallucinated the whole software!

rapnie• 3 years ago

4. Hallucinate the satisfied customer.

runnedrun• 3 years ago

I always assumed copilot and ai code assistants were replacing docs, for most use cases. I guess I'd rather just ask the question I have, when I have it, instead of having copilot preemptively create docs for questions it thinks I'll have later on.

siva72• 3 years ago

Ok, so why do we need stackoverflow anymore?

number62• 3 years ago

Internetpoints for the CV

blackoil• 3 years ago

To feed data to LLMs

figassis• 3 years ago

I don't think I've used it in ages. Maybe part of maturing as a developer, most answers I started to find there were expected and not covering the problems I faced, so usually original docs are the only reliable source, and maybe a GH issue to interact directly with the author. So I guess a tool to surface information straight from docs beats SO. Also won't complain about not having to deal with legitimate questions being closed due to "reasons".

siva72• 3 years ago

Same for me. The issue tracker is usually the most interesting place to figure out why it's not working as expected while in the beginning of my career it was oftentimes SO.

whoomp12341• 3 years ago

who else is going to tell me that all my ideas are stupid? an LLM?

guappa• 3 years ago

I think the next iteration of LLM will not be able to use newer libraries because of lack of training material on stackoverflow.

jorgehaddad07• 3 years ago

like overflowai

farnsworth• 3 years ago

If you're only using stackoverflow to have someone read the docs to you, then I feel you're missing the point. There is so much between the lines, and the LLM isn't 100% there yet.

mcny• 3 years ago

Is it possible to teach llm to read issues and answer questions such as how do I get angular 17 to output to dist instead of dist/browser with you can't do this with angular. However, in your CI CD pipeline, you can do a mv or robocopy or bla bla bla depending on context?

https://github.com/angular/angular-cli/issues/26028

It would be very useful because there are so many edge cases that aren't covered in the docs but are probably available in some kind of WON'T FIX issues.

monkeydust• 3 years ago

So another RAG product? I do like the ability to specify your knowledge and qualification level to help steer the output.

blueboo• 3 years ago

What is copilot for docs? This page, rather infuriatingly, doesn’t explain! Is it part of copilot? A product? An experiment…?

Copilot for docs is ___

pid-1• 3 years ago

Let me try:

GitHub Copilot for Docs™ is an award winning, Gartner® Magic Quadrant™ leading, blazing fast AI solution that empowers business leaders to maximize their ROI by leveraging state of the art GPU enabled Hybrid Cloud infrastructure on the Edge to break knowledge silos and improve productivity at scale.

bertil3• 3 years ago

An alternative to reading the docs; you can ask questions about the problem you are facing, and an LLM (presumably trained on coding questions, and RAGged to the docs) will give you configurable answers.

theonlybutlet• 3 years ago

Don't think they know yet either. Still early days, they're basically just dangling the carrot out of reach. They're GitHub researchers based on their Twitter.

dkdbejwi383• 3 years ago

So how do you actually use this? Or is it just a design/proposal?

bertil3• 3 years ago

Like many of those projects, access is controlled through a Discord server that is dedicated to more than just this prototype. I’m parsing and seeing things about getting CoPilot and VSCode to work together.

If you are eager to try something similar without dealing with a community, I’m assuming it’s possible to download the documentation from your favorite framework and start a “GPT” on OpenAI by giving it the role of a helpful coding assistant.

I definitely feel like there’s something to be done beyond the RAG aspect and figure out how to address the “other question” problem (you ask how you can do A, but that’s because A is your way to finding a solution for B, and there’s a much easier solution to B, but you don’t get the connection because you don’t understand the abstraction C). That would require a much more deliberate community to formalise properly.

ianbruton• 3 years ago

Got a link to the discord server?

nvm, I found it on the top right of the copilot for docs page

karmasimida• 3 years ago

Copilot for docs is a horrible name for this demo. It feels more like a search engine concept, the knowledge base is up to date, and the answer is personalized.

But again, the naming is horrible, I thought is was a Copilot just for markdown.

GuB-42• 3 years ago

I thought it was an autocompleter for writing documentation.

I am very disappointed that it isn't.

I had some success having ChatGPT document my code, it fact, it does a better job than I do by myself. Sometimes I have to fix some misunderstanding, but when it comes to writing and finding the right words, it is way better than I am. For me, it works better than for code generation, and I think that kind of "Copilot for Docs" would make a lot of sense.

bertil3• 3 years ago

This is an excellent idea, but it feels like an obvious application of RAGs (or a trivial agent to build with OpenAI’s latest ‘GPTs’ tool). Not dismissing the necessary and valuable work—it’s “obvious” in the way sleeping after a long day is obvious.

What I’d love to see is such an agent *learn* from the interactions with people asking:

* What questions they ask more often than other framework, and how can the documentation clarify that (explicitly or not);

* Can you detect when questions betray a naive grasp of the problem and when the person asking would benefit from dedicated training, not just a quick answer;

* Can an LLM structure programming concepts and suggest ways to describe a framework that would help people make sense of what each framework actually is?

infecto• 3 years ago

Who made this website? I feel like I am in a vegas casino and having trouble focusing.

It was unclear in the quick pitch but I am wondering if they essentially throwing RAG/search on top of a GPT model? I am guessing so because I cannot imagine you could train the model on such a limited source and get meaningful information. If thats the case perhaps this will be interesting but I think there are other interesting angles to this approach than focusing on a single library/codebase.

orbz• 3 years ago

That’s an official GitHub website surprisingly.

scottmf• 3 years ago

I switched to Cursor (VSCode fork) a couple of months ago. It does this already and I haven’t looked back since. Especially as the copilot extension (along with all my others) work out of the box.

It has full knowledge of your entire codebase rather than being limited to currently open files (I assume embeddings with RAG) and will index documentation (or any other URL).

I find it especially useful for things like Swift/SwiftUI APIs given how quickly they can evolve, not to mention copilot’s other limitations when it comes to languages beyond JS/TS, Python, etc

canadiantim• 3 years ago

How can you be sure Cursor isn’t just pilfering your entire codebase? Where’s the trust factor? Does look cool tho

jerrygoyal• 3 years ago

copilot recently added feature to "ask your codebase"

agloe_dreams• 3 years ago

I unironicly love the sliders below the search box.

toddmorey• 3 years ago

Hate how they announce this product by shitting on the hard work of docs teams.

I get that docs are uneven in quality but there’s also been what I think is a renewed focus on quality & interactive examples. There’s some truly fantastic OSS documentation out there.

The tone of this just didn’t sit right with me.

dvt• 3 years ago

Love how the AI hype went from "it's going to write code for you!" to "oh okay, maybe just boring boilerplate" to "it's basically a glorified search" (and it probably won't be great at that, either).

danielmarkbruce3• 3 years ago

Lot's of people (including me) use co-pilot and chat gpt to save several hours a week, likely > 1 hour a day. It's under hyped.

sdwr• 3 years ago

Yeah, it saves me time + effort, and lets me reach for larger changes than I otherwise would. It acts as a safety net / scaffolding as well as just answering Qs / writing code.

mpalmer• 3 years ago

Counter-hype is becoming just about as tiresome as hype. Have you used AI in anger for any of these use cases?

gettothegold• 3 years ago

This is a great use of AI

mooreds• 3 years ago

We're using kapa.ai for this right now, pretty happy with them.

It's an LLM trained on our docs, openapi spec, github issues, and forum posts.

Definitely been helpful to our customers and users to find answers/pointers to docs more quickly.

whalesalad• 3 years ago

Is there a way to do the inverse: produce a rough draft of a complete documentation site based on reading a codebase?

Copilot for Docs

Related Stories