Chroma

Database execution engines are an incredibly fun theoretical computer science and practical engineering problem to work on. I wrote about how Sanket Kedia Sicheng Pan and I designed Chroma's execution engine. https://lnkd.in/gzprXKBF

Are you using Chroma in Prod? Send us a DM for our new swag!

We're building a horizontally scalable, cloud-native distributed system, designed from the ground up to power retrieval for AI application workloads. Designing a serverless retrieval system is challenging; there are many trade-offs and constraints unique to retrieval for AI applications which must be carefully considered. In particular, retrieval workloads for AI applications differ significantly from traditional search workloads. Everything must be designed with the requirements of AI applications in mind. Learn about how we're using object storage to tackle this challenge here: https://lnkd.in/gu3Sa8hv

Introducing AI Dev Explainer (https://lnkd.in/gRjirVUt) - The best resource for getting started with building AI applications with LLM. We built this explainer because the most important thing anyone can be doing in AI today is onboarding more developers onto building with it. AI is great but it has often been marketed remarkably poorly. 'Generative AI' makes it sound like something that writes college essays, and most AI discussion makes it seem like you need a PhD. to understand it. But it's just an API anyone can use. We looked around at what's around on the web and found that resources are very fragmented, used a lot of jargon, and mainly focused on on people already building with AI. Our aim is to get the next million developers started. Part one is out now. Part two is coming soon. We welcome feedback, and will be iterating on this constantly and adding more topics, in response to the community. Share with your developer friends just getting started in AI!

NEW: Rahul Sonwalkar on building Julius AI https://lnkd.in/g2QJxWJC

we're hiring a community lead at chroma if you are passionate about open-source and accelerating AI engineering, apply! https://lnkd.in/dBePEexU

Embedding Adapters Today, we are pleased to share the first of a series of technical reports with the AI application developer community—our investigation into the use of linear embedding adapters in improving retrieval accuracy in realistic settings. Retrieval accuracy is an important determinant of AI application performance. However, many approaches to improving retrieval accuracy require large labeled corpora, which are often not available to application developers. Additionally, many of these approaches require re-computing the entire set of embeddings. While embedding adapters aren't a new idea, but to our knowledge this is the first time they have been investigated in depth. In this work, we demonstrate that applying a linear transform, trained from relatively few labeled data points, to just the query embedding, produces a significant (up to 70%) improvement in retrieval accuracy across many domains, including across languages. For many applications, this is the difference between working or not. Learn more: https://lnkd.in/gnC_FTFm

We are hiring for a founding talent / recruiter

By popular demand, Chroma is supporting Backdrop Build V3. We have grants available for excellent projects in the following categories: Autonomous AI systems - Programs backed by large language models (LLMs) which are capable of operating on their own, without direct human prompting, to run a process, gather information, or accomplish a goal. They might use retrieval for their working memory. Systems that learn - Programs which improve and adapt either implicitly through use, or explicitly by being taught and then remembering skills they've learned at the right time. They might use retrieval to store and remember skills and tools. Systems that explore - Programs that are capable of recognizing when knowledge is not yet available, and either ask for help or autonomously seek out relevant new knowledge as needed. They might use retrieval to figure out what they don't yet know, and where to find out. Multi-modal systems - LLMs are now capable of processing not just text but also images, and audio. Chroma supports multi-modal retrieval, and we want to see exciting applications of these capabilities. We are interested in effective demonstrations of these capabilities, far more than full-fledged products. Take Voyager as your inspiration: https://lnkd.in/eKU_f8UF Look into the future of what might be possible, even as a toy, rather than what might be commercially viable today. Apply here: https://backdropbuild.com/

It’s been a year since we launched Chroma's open source embeddings store, making it easy to build retrieval into your AI application. Since then, over 7.3 million individual machines have run Chroma. The ecosystem has evolved, and we’ve learned a lot. Over the last twelve months, our thesis that retrieval would be a fundamental component of AI application development has paid off. Retrieval is also becoming an increasingly important component in AI research. Our investment in developer experience and usability, as well as our integrations with other open-source projects like LlamaIndex and LangChain, have earned our leading position in retrieval in Python, and in open source overall. So, what's next? Everybody already knows Chroma's developer cloud is coming, and soon. At launch, it will have the best product, with the best developer experience, at the best price point. Yes, it's taking longer than expected - we are confident the greater investment is worth it. Almost since launch, we've been saying that vector search alone isn't enough. The last year has shown more of what works and what doesn't in retrieval for AI. It was gratifying to see OpenAIDevs go into detail about what it takes to build retrieval that actually works at developer day.(https://lnkd.in/gBWgYmY3) Automatic embedding model selection, and the dual problem of optimal chunking, result relevancy and ranking, and automatic fine-tuning of the retrieval system aren't optional extras, but critical components which we'll be building into our product. Neural retrieval approaches, like @lateinteraction's CoLBERT, show a lot of promise - it's a natural extension of both retrieval and language models to use much more of the context of both query and document when retrieving. Looking further ahead, the retrieval-AI loop ('RAG') as it's done today is very primitive. Stuffing the LLM's context window with data blindly retrieved from elsewhere, and mixing instruction, data, and other context together is clearly suboptimal. Over the next 12 months, Chroma will be experimenting with architectures and approaches that more directly integrate the retrieval system directly with the execution LLM. There are a lot of promising directions, and open-weights models will help a lot. It's very clear that it's still very early in AI. I often draw an analogy to the early web, or to the early days of aviation - the best thing people can be doing right now is experimenting. Chroma will continue to make it as easy as possible to experiment with AI. LFG.