Originate HN: Sidekick (YC W23) – AI red meat up bot for developers
Hi Hacker Info! We’re Jason and Ayan, the cofounders of Sidekick (https://www.getsidekick.ai/). We made a GPT-pushed bot for developers that runs in Slack and Discord and answers questions about your developer doctors, while robotically conserving them up so some distance with contemporary info. You must well maybe also be part of our public Slack the set aside you must be ready to quiz Sidekick questions about Airbyte, a fashionable originate supply records connector catalog:
https://be part of.slack.com/t/sidekick-public/shared_invite/zt-1r…
Or ought to you rob to no longer test in to our Slack, right here’s a demo video showing the same factor: https://youtu.be/fPhP1325RkI
We’re one day of of developing every little thing originate supply (there are some contractual disorders we’re working by), nonetheless our client facet code and fundamental infra is right here: https://github.com/ai-sidekick/sidekick.
Providing technical red meat up to developers has been costly for companies because they need to rent professional engineers to attain it. We’ve viewed community red meat up channels with a 2000:1 ratio of developers to red meat up engineers – there’s no manner every demand of will derive answered. We constructed Sidekick to invent this extra special easier. It’s seriously helpful for originate-supply companies/projects because many OSS communities have pretty about a of us asking questions, nonetheless no longer steadily someone helping troubleshoot.
We integrate with Slack and Discord, since that’s the set aside developer red meat up is already going on. On the backend, we teach Weaviate to index the records and OpenAI’s text-davinci-3 model to generate responses.
Moreover to answering questions, Sidekick could well maybe also also update .md files robotically with contemporary info. When somebody reacts to a message in Slack with the emoji, Sidekick will teach Weaviate to search out the section for your documentation that’s most connected to the message, then teach GPT3 to merge the contemporary info into the documentation. At closing, this will also simply submit a pull quiz on Github with the adjustments. We saw that devrel groups are already making product announcements and helping customers troubleshoot unusual disorders in the community, so we constructed this selection to keep them design more time.
We teach GPT for producing the responses and contemporary documentation, nonetheless are relying less and less on it after discovering out that you simply hit a ceiling on acknowledge quality very rapid by utilizing handiest GPT and in fact helpful engineering ways. Right here’s some of what we learned looking out to forestall hallucinations in our answers: https://medium.com/@jfan001/how-we-slit-the-rate-of-gpt-hallu…
What we learned makes a extra special bigger difference is the breadth and quality of the boom material you must be ready to search by, which is why we now rely design more on cleaning and annotating records, which yields seriously better outcomes when blended with in fact helpful chaining. As an instance, as an different of naively chunking records into 1000 token blocks, we parse the markdown into semantically meaningful sections (e.g. paragraphs, lists, code blocks) and ticket the boom material with the header title and file title so it’s prone to floor for searches which can be match for the piece it’s from, even supposing it doesn’t precisely match the boom material in that chunk.
One relaxing factor we also learned is that when Sidekick gets added to a #serve channel, of us that otherwise wouldn’t quiz questions delivery using it. It turns out, there are pretty about a “lurkers” who arrive to those channels to search out answers, nonetheless don’t are looking out to bother someone with their distress. Adding a tool that they’ll derive answers from straight away brings these of us out into the community, giving founders and community managers a likelihood to effect out to them.
To summarize, Sidekick 1) saves red meat up engineers time, 2) retains the doctors up so some distance and 3) helps hold developers in the community. Lengthy-length of time we’re looking out to produce an analytics product on top of Sidekick so companies can realize how their product is being inclined and the set aside there are alternatives to add more price to their possibilities (and rate more cash for it).
We’d like to listen to from the HN community about this product! Get you utter using a tool to search by and update developer doctors from Slack would keep you time?