kapa.ai

"The new image indexing feels like magic." Last week I posted about the least glamorous kapa.ai feature we shipped this year: showing images in answers. It hit the front page of Hacker News for the complex retrieval engineering behind it. But this is the part I actually care about. An email like this from a customer, unprompted, a week after rollout. Their docs are heavily visual, like most software products. And their users are noticing the AI "getting better and better" without knowing why. A/B tests told us this feature worked but emails like this tell us it matters. :)

AI adoption inside a company is never a wave. After 200+ agent rollouts since YC, I can tell you exactly how it happens: three people go all-in while everyone else watches 🧵 Quick context: kapa.ai is how teams like Mixpanel, Sentry, and Grafana build insanely accurate chat on their technical content - support engineers resolving tickets, AEs answering deep technical questions mid-deal, solutions teams filling RFPs without pinging engineering. Internally it's always the same pattern: a few people build kapa into their daily workflow immediately, most follow over weeks, some never touch it. The deployments that take off? The teams that found those early power users and let them train everyone else. The champions had already figured out the workflows - they just needed a spotlight. But spotting them is getting harder. Usage is spread across Slack bots, MCP servers, and chat. Nobody has the full picture. So we shipped Top Users for internal kapa projects: 1/ See who depends on kapa most across every surface, with questions asked and time saved 2/ Spot which roles get the most value, so you know where to expand next 3/ Turn "our team uses AI" into "this AE self-serves 90% of technical questions" (an actual ROI story!!) Fun fact from our own dashboard: kapa's heaviest internal users are our AEs. :)

Best Slack message I've received this year. A friend at OpenAI messaged me from a conference: a unicorn founder on stage just described their AI support agent journey - they tried multiple solutions that all "ended up being very sh*t"... and then named kapa.ai as the one they finally stuck with. I'll take that over any award 😄 There's a real lesson in it though. Everyone's AI support demo looks amazing. The gap between tools only shows up in production, months in, when real users ask real questions and either trust the answers or quietly stop using it. That's the whole game: be the tool still standing after the "very sh*t" phase of everyone else. No case study, no logo slide, no marketing budget can buy an unprompted mention on a conference stage. Screenshot in the post because moments like these are why we grind on accuracy 🔥

We just made the front page of Hacker News. For the most obvious feature we shipped this year: showing images in answers. Turns out obvious ≠ easy. Why obvious? Technical docs were never just text. Wiring diagrams, schematics, spec tables - sometimes the honest answer is "look at this diagram". Yet most retrieval systems throw all of it away at ingestion and answer with words alone. Why not easy? The naive approaches don't survive contact with production: → Multimodal embeddings wash out the fine detail that matters in charts and schematics → Sending images at query time adds 27-51% cost per query and blows past payload limits Our approach: describe every image once at ingestion, retrieve as text. Per-query overhead drops to 1-6%, and that engineering story is what HN picked up. And the impact? We A/B tested across half a million real user interactions on kapa.ai - it lifted CSAT more than any LLM model swap in the past year. Rolling out to all kapa deployments now. Hardware folks with diagram-heavy docs, this one's for you. Sometimes the answer isn't a smarter model. It's showing users what they asked to see. :) Blog + HN thread in the comments 👇

kapa now lives inside Zendesk too. 🎉 Your technical support agents don't leave the ticket. kapa reads the full thread, drafts a reply grounded in your docs, and stages it in the composer for one-click insert. The agent reviews and hits Submit. It also summarizes any ticket for clean handoff, and answers your agents' own questions on the side. Same kapa accuracy 200+ technical teams trust, now right where your support team already works. Live now 👇

Your technical support team keeps solving the same ticket twice. Not because they're careless - because the fix from six months ago lives in a closed ticket nobody will ever find. We just shipped Salesforce Cases as a data source for kapa.ai - the last big ticketing system after Zendesk and Jira Service Management. Why? Your case history is verified problem-and-solution pairs written by your best support engineers. Docs tell you how the product should work. Tickets tell you how it actually breaks, and how to fix it. The most common ways teams use it: 1/ Surface similar cases and past resolutions the moment a new ticket comes in 2/ Give L1 support answers that used to require an escalation 3/ Spot recurring issues that reveal exactly what your docs are missing 4/ Build your own support automations (auto-triage, agent-assist, account summaries) with kapa as the grounded retrieval layer Your support history is a goldmine. Stop treating it like an archive. Docs in the comments 👇