Loading video...

Video Failed to Load

Go Home

🦆🦆🦆 GPTDuck - question answering against any Github repository. Simply type in a Github repo, and then ask questions about the repo. Here's me asking a questiont about how GPT News works. Try it out here: or in Slack via

333,664 views • 3 years ago •via X (Twitter)

10 Comments

Gary Wu's profile picture
Gary Wu3 years ago

Biggest limitation right now is repo size - only supports repos smaller than 100MB. Support for private repos is in beta - retweet and comment for free access. Am playing with code generation as well, but results are a bit finicky. Feel free to DM me if any issues / questions.

Gary Wu's profile picture
Gary Wu3 years ago

Built with @langchain and @OpenAI embeddings. The name comes from 'rubber duck debugging' - this is an AI duck that is hopefully more helpful.

Gary Wu's profile picture
Gary Wu3 years ago

ok I've been GitHub rate limited - fixing now

Gary Wu's profile picture
Gary Wu3 years ago

Rate limits fixed - sign up and add your own Github access token to get around 500 errors. Now my @supabase instance is struggling to stay alive💀

Jerry Liu's profile picture
Jerry Liu3 years ago

@gpt_index Nice! Do you just scrape the README or all files within the repo? Depending on which queries you'd want to support, I wonder if @gpt_index could help you synthesize some of these answers through an index struct

Gary Wu's profile picture
Gary Wu3 years ago

@gpt_index It indexes the entire repo! So you have access to functions as well if needed. Absolutely - want to add it for larger repos especially.

Harrison Chase's profile picture
Harrison Chase3 years ago

@gpt_index very cool! are you using some sort of vector db + semantic search under the hood?

Gary Wu's profile picture
Gary Wu3 years ago

@gpt_index yup, exactly that!

Charly Wargnier's profile picture
Charly Wargnier3 years ago

@gpt_index Great app! I've got this issue, is this expected?

James M's profile picture
James M3 years ago

@gpt_index Love it, I've built something similar that indexes Confluence Spaces and Zendesk Academy (but only in memory) Are you taking all the code in chunks of up to 2046 tokens or splitting down by file/function?

Related Videos