- Chroma helps manage vector embeddings, or numerical representations of data AI models can process.
- This is a key technology for generative AI founders building on top of models like OpenAI’s GPT-4.
- The startup raised $18 million in seed funding at a $75 million valuation led by Quiet Capital.
In 2018, when Anton Troynikov finished graduate school in Germany and moved back to the US, he was itching to dive back into the Bay Area’s tech scene.
He immediately emailed every computer vision or robotics founder from the previous two Y Combinator batches at the time and started a budding friendship with entrepreneur Jeff Huber.
Now, five years later, the two are teaming up to build their own startup: open-source vector database Chroma. The company recently closed a $18 million seed round led by Quiet Capital at a $75 million valuation. Angel investors like Angelist cofounder Naval Ravikant, OpenAI founder Sam Altman’s brothers, Max and Jack Altman, Vercel CEO Guillermo Rauch, and MotherDuck CEO Jordan Tigani also contributed. This follows a $2.3 million pre-seed round raised from AIX Ventures, Bloomberg Beta, AI Grant, and Replit CEO Amjad Masad last year, the two cofounders told Insider.
Originally, Chroma’s tech aimed to serve machine learning engineers who were training and deploying AI models. However, when OpenAI’s viral ChatGPT came out late last year, the two realized that the number of developers building on top of these models was poised to skyrocket and shifted the company’s focus
Popular AI models like OpenAI’s GPT-4 are powerful general purpose models. However, there is a lot of data that doesn’t exist in its training sets like private company-specific data, specialized data not broadly available, or data that emerged after it was trained.
Chroma helps developers leverage this data by allowing them to plug into models vector embeddings, or representations of unstructured data that describe them as numeric values in hundreds or thousands of different dimensions, enabling data to be grouped together based on similarity.
For instance, one common use case appears in developers looking to create a “ChatGPT” for their proprietary data or documents. Using Chroma, users can embed those documents into vectors and easily ask questions about their data. Because this limits the model’s context window, this approach tends to prevent hallucinations, or confident but incorrect responses from AI, Troynikov told Insider.
Tools like Chroma may become more popular as founders look to build domain-specific startups on proprietary data to differentiate from the rapidly-growing expanse of generative AI upstarts.
Older vector database startups, like Pinecone and Weaviate, have been around for years. However, these companies are more targeted at infrastructure engineers versus developers building AI apps, Troynikov said. These disparate focuses require different capabilities. For instance, when conducting semantic search, or searching based on meaning and content, an infrastructure engineer searches billions of mostly static entries in one flat index, while an app developer searches many dynamic smaller indexes, one per user, Troynikov explained.
Additionally, Chroma differentiates through its open source approach and scalable pricing, Huber said. And within the developer community, the startup is known for its ease-of-use, a tool that’s “much easier to get going if you want to use LLMs,” Essence VC managing partner Tim Chen told Insider.
Following its seed round, Chroma is aiming to release the hosted, paid version of its product within the next two to three months and hire engineers and applied researchers, Troynikov said.
The two cofounders say they’re excited about the rapid pace of advancements in AI and how that mirrors the endless possibilities of earlier tech eras.
“If you logged into Yahoo in 1995, there’s no way for you to predict Snapchat,” Troynikov said.
Read the full article here
