VecML’s mission is to build the next-generation AI infrastructure to facilitate AI Agents and many different Large Language Model (LLM) applications, including personal assistants, smart RPA, AI-enabled phones, recommendation/advertising systems, AI-assisted medical diagnosis, humanoid robots, etc. We have released the initial trial version of our SaaS products for neural search engine, privacy-protection engine, statistical modeling, machine learning (ML) platform, and Retrieval Augmented Generation (RAG) platform.
The founders, core engineering team, and advisory committee of VecML consist of renowned scientists and highly experienced engineers in AI, machine learning, search, databases, knowledge graphs, privacy, and edge computing. Dr. Ping Li, the CEO, has been ranked highly by www.csrankings.org (score 44.6 listed by Rutgers University).
Take the recommendation/advertising system as an example, which is the major cash cow for many IT companies. One million (or one billion) products are represented as (feature) vectors through LLMs or special purpose embedding algorithms. These one million (or one billion) vectors are stored in the “neural search engine”. If privacy protection is needed (e.g., when they are stored in the public cloud), these vectors are first processed via our novel privacy-protection engine before they are stored in the neural search engine. When a customer query arrives, a query embedding is generated and compared with the one million (or one billion) vectors in the neural search engine. The best 1000 products (vectors) are selected as the candidates which are re-ranked by the ML platform (typically a neural network) to produce the top-3 or top-10 recommendations.
In summary, VecML offers a variety of products for AI infrastructures including:
- · Neural search engine: It differs from the usual “vector databases” in that we allow vector similarities to be computed from a neural net. The paradigm of “fast neural ranking” is the revolution in approximate near neighbor search. Other key features include search with constrained (business filters), GPU for fast search, search with privacy, search with sparse data, search integrated with machine learning, etc.
- · Privacy engine: Embeddings leak private information of the original data. When embeddings are stored in the public cloud, or used for federated learning, privacy is a serious concern. We have resolved some of the most challenging problems in privacy protection. With our products, customers can enjoy private search and machine learning without sacrificing speed/accuracy.
- · Statistical ML platform: With our platform, machine learning (ML) has become light-weight and more accurate, very suitable for scenarios (such as edge computing) where resources are restricted. Our ML platform is built on top of the neural search engine. This is very convenient, because otherwise engineers have to first store the vectors in the vector database and then copy the vectors to another platform for training ML models.
- · Embedding generation engine: The quality and accuracy of LLM applications, to a large extent, depends on the quality of embeddings, which are representations of objects.
- · RAG, Knowledge Graphs (KG), prompt engineering engine: The retrieval augmented generation (RAG) has become the standard approach for effectively utilizing private/updated data in LLM applications. The private data are typically represented by embedding vectors or knowledge graphs (KG). The team is experienced with both Vector-RAG and KG-RAG. Some team members developed the world-first HNSW-GPU platform and some developed the world-first supervised neural Open-Domain Information Extraction (OIE) platform.