PatentPT: Patent search with LLM-powered solutions

In this post, we unpack another insight from the YOLO VISION 2023 (YV23) event, hosted by Ultralytics at Google for Startups Campus in Madrid.

In this blog we’ll explore the talk given by the founder of Activeloop, Davit Buniatyan, as he takes us through the genesis of PatentPT, an LLM-powered solution for patent search.

Unveiling PatentPT#

Ever felt overwhelmed by the sheer volume of patent data and the tedious search process? PatentPT was built to address exactly that.

Led by Davit Buniatyan, this talk shares practical insights into fine-tuning and deploying Large Language Models (LLMs) for patent autocompletion, abstract and claim generation, and advanced search functions within a rich patent corpus.

Activeloop and DeepLake: A Unified Data Storage Layer for AI#

Before we dive into the nitty-gritty of PatentPT, let's take a look at Activeloop's creation: DeepLake, a database for AI. With the AI data stack fragmented across various storage systems, DeepLake offers a unified data storage layer that streamlines AI workflows.

From storing metadata to unstructured data and embeddings, DeepLake simplifies the process, so data scientists can focus on training ML models without the hassle of data management.

Exploring DeepLake's architecture and features#

Now, let's delve into DeepLake's architecture and features. With its open-source components and serverless design, DeepLake enables data storage and versioning on object storage and connects directly to ML models. It also features Deep Memory, a feature which enhances search accuracy without altering embeddings.

Deep memory demo: Elevating patent search#

Davit allowed us to delve deeper into this workflow with a live demo showing how Deep Memory improves patent search. We got a firsthand insight into how Deep Memory delivers up to 22% accuracy improvement with sub-second queries at a fraction of the cost compared to traditional solutions.

The result is faster, more precise patent search without endless scrolling through patent databases.

The genesis of PatentPT: From concept to reality#

Ever wondered how PatentPT came into existence? Let's rewind the clock and take a closer look at the comprehensive steps taken to create this solution. From LLM model training and fine-tuning to crafting custom features and deploying search APIs, Davit Buniatyan walked through each step the Activeloop team took.

Unleashing the power of LLMs: The future of AI solutions#

Overall, PatentPT exemplifies the potential of LLM-powered solutions in specialized fields like patent search. Activeloop's work on DeepLake points toward AI solutions that make unstructured data faster and cheaper to use.

Wrapping up#

As AI continues to advance, the real value lies not just in the technology itself, but in how it helps us solve real-world challenges. Join our community, check out our docs and our GitHub repository to stay updated on the latest advancements!