Generative AI and large-scale language models (LLM) have been all the rage in recent years, upending traditional online search with the likes of ChatGPT while improving customer support, content generation, translation, and more. Now, a fledgling startup is using LLM to build an AI assistant that can answer complex questions from developers, software end users, and employees, among other things. It is similar to ChatGPT, but targeted towards technical products.
Founded last February, Kapa.ai is an alumnus of Y Combinator's (YC) Summer 2023 program and already has a pretty impressive roster of customers, including ChatGPT creator OpenAI, Docker, Reddit, Monday.com, and Mapbox. are collecting. Not bad for an 18-month-old business.
“Our initial concept came after several friends who run technology companies consulted about the same problem, and we created the first prototype of Kapa.ai to address this problem for them. After building it, we ran our first paid pilot within a week,” founder Emil Sorensen told TechCrunch. “This has led to natural growth through word of mouth. Our customers have become our biggest supporters.”
Building on its early traction, Kapa.ai has now raised $3.2 million in a seed round of funding led by Initialized Capital.
learn technical things
In the broadest sense, companies enter their technical documentation into Kapa.ai and provide an interface for developers and end users to ask questions. For example, Docker recently released a new documentation assistant called Docker Docs AI. It instantly answers Docker-related questions from within the documentation page. It is built using Kapa.ai.
Kapa.ai on Docker. Image credit: Kapa.ai
But Kapa.ai can be used for countless use cases, including customer support, community engagement, and workplace assistants that allow employees to query a company's knowledge base.
Under the hood, Kapa.ai is based on multiple LLMs from different providers and leverages a machine learning framework called Search Augmentation and Generation (RAG). This improves LLM performance by providing richer data that can be easily retrieved from relevant external data sources. response.
“We are model agnostic. We work with multiple providers, including using our own models, to use the best performing stack and acquisition technology for each specific use case.” said Sorensen.
Note that there are already a number of similar tools, including from venture-backed startups like Sana and Kore.ai, that are essentially aimed at bringing conversational AI into enterprise knowledge bases. Worth it. Kapa.ai also fits into that category, but the company says its main differentiator is its primary focus on external users rather than employees, which has heavily influenced its design. That's what it says.
“When you bring an AI assistant to an external end user, the level of oversight increases tenfold,” Sorensen says. “Accuracy is all that matters. Businesses are worried about AI misleading customers, and we've all tried ChatGPT or Claude to hallucinate. A few wrong answers. Companies quickly lose trust in the system, so that's what we care about.”
accuracy
This focus on providing accurate responses on technical documentation while minimizing hallucinations highlights how Kapa.ai is a different kind of LLM animal. That means it's built for a much narrower use case.
“Optimizing a system for increased accuracy naturally involves a trade-off: the system must be designed to be less creative than other LLM systems would tolerate. ,” Sorensen said. “This is to ensure that answers are generated only from the world of content provided.”
Then there's the thorny issue of data privacy. This is one of the major deterrents for companies considering implementing generative AI but wary of exposing sensitive data to third-party systems. That's why Kapa.ai includes PII (Personally Identifiable Information) data detection and masking to ensure that your personal information is not stored or shared.
This includes real-time PII scanning. When Kapa.ai receives a message, it scans the PII data and if personal data is detected, the message is rejected and not stored. Users can also configure Kapa.ai to anonymize PII data found within documents.
Of course, companies can assemble something similar to Kapa.ai themselves using third-party tools like Azure's OpenAI service or Deepset's Haystack. But this takes time and resources, especially when you can tap into Kapa's website widget, deploy its bots to Slack or Zendesk, and use APIs that allow businesses to customize their own interfaces a bit. It's a consuming endeavor.
“Most of the people we work with either don't want to do all the engineering work or don't necessarily have the AI resources on their team to do the engineering work,” Sorensen says. . “They want accurate and reliable AI engines that are reliable enough to expose directly to customers and that are already optimized for the use case of answering technical questions for their products.”
Regarding pricing, Kapa.ai says it has a SaaS subscription model and offers tiered pricing based on the complexity of deployment and usage, but it does not disclose these prices. .
The company has a remote team of nine people around the world, with two main locations in Copenhagen, where Sorensen is based, and San Francisco.
In addition to lead backer Initialized Capital, Kapa.ai's seed round also includes Y Combinator, Docker founder Solomon Hykes, Stanford University professor and AI researcher Douwe Kiela, and Replit founder Amjadmasad. Many angel investors participated, including Mr.