January 2024 Summaries

The blogpost discusses the technical journey behind building a fully custom LLM-based retrieval augmented generation and search app, PatentPT. It highlights the features of PatentPT, its technical architecture, dataset creation, domain training, finetuning large language models for patent generation, creating custom featurizers, standing up search indices, deploying search APIs, deploying LLM inference APIs, and the final application. The post emphasizes that while the stack for training and deploying fine-tuned LLMs is not yet solidified, PatentPT showcases an efficient approach using cutting-edge technologies like Deep Lake from Activeloop, Hugging Face Optimum by Intel, and Habana Gaudi HPU hardware.

The blog post details the development of a custom application named PatentPT, which leverages fine-tuned large language models (LLMs) to enhance patent search and retrieval processes. In response to the need for more practical guides on deploying LLMs on custom text corpuses, the authors describe how they domain-trained and fine-tuned an LLM using the vast United States Patent and Trademark Office (USPTO) dataset. The application, PatentPT, improves upon traditional patent search methods by providing features such as autocomplete, abstract and claim generation, and general chat functionalities. The technical architecture involves creating an ensemble of fine-tuned LLMs and custom search indices, utilizing tools like Habana Gaudi hardware, Deep Lake databases for AI, and Hugging Face's Optimum library to achieve efficient training and deployment. The result is a scalable, LLM-powered application that demonstrates greater accuracy and control over output than general AI APIs, showcasing the integration of cutting-edge technologies in the field of large language modeling.

Activeloop has achieved SOC 2 Type 2 certification, demonstrating its commitment to data security and operational excellence in the software-as-a-service sector. The company's rigorous standards for safeguarding customer information align with five key trust service principles: security, availability, processing integrity, confidentiality, and privacy. This certification is a testament to Activeloop's dedication to maintaining high data security standards for its customers and users, ensuring ongoing compliance and reliability. The process of achieving this certification involved a comprehensive audit by an independent auditor and a documentation process, validating the company's practices in data security and operational integrity.

Vector databases like DeepLake form the foundation for many AI applications. They provide the capacity to store and retrieve complex, high-dimensional data, enabling functionalities like Retrieval Augmented Generation (RAG) and sophisticated recommendation systems. Alongside vector databases, Large Language Model (LLM) frameworks such as LlamaIndex and LangChain have emerged as key players in accelerating AI development. By simplifying the prototyping process and reducing development overheads associated with API interactions and data formatting, these frameworks allow creators to focus on innovation rather than the intricacies of implementation. In this blog post, we walk you through constructing a complex and interactive shopping assistant using DeepLake and LlamaIndex. This assistant exemplifies how intelligent systems can be built from fundamental components like vector databases and LLMs. The project is an AI-powered shopping assistant designed to leverage image processing and LLM agents for outfit recommendations, providing tailored outfit suggestions based on user input. The architecture design of the application follows a linear yet dynamic flow, with the LLM agent at the helm. Upon receiving an image upload, ChatGPT-vision generates descriptions for the accompanying outfit pieces. These descriptions guide subsequent searches in DeepLake's vector database, where the most relevant items are retrieved for each piece. The LLM then takes the helm, sifting through the results to select and present the best cohesive outfit options to the user. The shopping assistant is designed to deliver not only outfit suggestions but actionable shopping options, providing real product IDs (that can be converted into URLs to retailers) along with price comparisons. Throughout the course, you will learn how to extend the AI's capabilities to facilitate an end-to-end shopping experience. The user interface of this application is designed with functionality and educational value in mind. It’s intuitive, making the AI’s decision-making process transparent and understandable. You’ll interact with various application elements, gaining insight into the inner workings of vector databases and LLMs. In conclusion, the roles of DeepLake and LlamaIndex have proven pivotal in developing an AI-powered shopping assistant. Their flexibility and power could drive innovation across various domains, from healthcare to finance and education to creative industries. Your insights and feedback are crucial as we continue to navigate and expand the frontiers of artificial intelligence.