Open HN: Sidekick (YC W23) – AI enhance bot for builders
Good day Hacker News! We’re Jason and Ayan, the cofounders of Sidekick (https://www.getsidekick.ai/). We made a GPT-driven bot for builders that runs in Slack and Discord and solutions questions about your developer docs, while mechanically retaining them up to this point with unique data. That you would possibly presumably also join our public Slack where that you would possibly want to inquire Sidekick questions about Airbyte, a customary originate offer knowledge connector catalog:
https://join.slack.com/t/sidekick-public/shared_invite/zt-1r…
Or at the same time as you bought to no longer signal up to our Slack, that is a demo video exhibiting the identical component: https://youtu.be/fPhP1325RkI
We’re within the strategy of making all the pieces originate offer (there are some contractual concerns we’re working through), however our client facet code and customary infra is here: https://github.com/ai-sidekick/sidekick.
Providing technical enhance to builders has been costly for companies because they desire to rent skilled engineers to lift out it. We’ve considered community enhance channels with a 2000:1 ratio of builders to enhance engineers – there’s no manner each ask will procure answered. We constructed Sidekick to derive this noteworthy more uncomplicated. It’s in particular helpful for originate-offer companies/projects because many OSS communities hold quite loads of folks asking questions, however no longer often any individual serving to troubleshoot.
We integrate with Slack and Discord, since that’s where developer enhance is already happening. On the backend, we utilize Weaviate to index the records and OpenAI’s textual order material-davinci-3 mannequin to generate responses.
Along with answering questions, Sidekick can also update .md recordsdata mechanically with unique knowledge. When anyone reacts to a message in Slack with the emoji, Sidekick will utilize Weaviate to search out the piece for your documentation that’s most linked to the message, then utilize GPT3 to merge the unique data into the documentation. At most life like doubtless, this could well submit a pull query on Github with the adjustments. We saw that devrel groups are already making product bulletins and serving to customers troubleshoot customary concerns within the community, so we constructed this characteristic to set them noteworthy more time.
We utilize GPT for generating the responses and unique documentation, however are relying less and less on it after studying that you hit a ceiling on answer quality very fleet by the utilize of easiest GPT and instantaneous engineering methods. Right here’s about a of what we realized searching for to forestall hallucinations in our solutions: https://medium.com/@jfan001/how-we-scale again-the-rate-of-gpt-hallu…
What we chanced on makes a noteworthy greater inequity is the breadth and quality of the order material that you would possibly want to search through, which is why we now rely plenty more on cleaning and annotating knowledge, which yields a ways better outcomes when mixed with instantaneous chaining. To illustrate, in desire to naively chunking knowledge into 1000 token blocks, we parse the markdown into semantically main sections (e.g. paragraphs, lists, code blocks) and mark the order material with the header name and document name so it’s inclined to floor for searches which can also very effectively be match for the piece it’s from, although it doesn’t precisely match the order material in that chunk.
One fun component we also realized is that when Sidekick will get added to a #encourage channel, folks that in every other case wouldn’t inquire questions originate the utilize of it. It turns out, there are quite loads of “lurkers” who come to these channels to search out solutions, however don’t desire to effort any individual with their teach. Adding a tool that they would possibly be able to procure solutions from straight away brings these folks out into the community, giving founders and community managers a possibility to set apart out to them.
To summarize, Sidekick 1) saves enhance engineers time, 2) keeps the docs up to this point and 3) helps do away with builders within the community. Long-term we desire to manufacture an analytics product on high of Sidekick so companies can mark how their product is being inclined and where there are alternatives to be capable to add more ticket to their potentialities (and fee more cash for it).
We’d like to hear from the HN community about this product! Homicide you judge the utilize of a tool to search through and update developer docs from Slack would set you time?