Harvey has unveiled a new agentic retrieval system designed to address the complexities of patent research. Unlike traditional retrieval-augmented generation (RAG) systems, which rely on fixed indexes and predictable research paths, Harvey's approach dynamically adapts to the evolving needs of each query. This innovation aims to streamline the intricate process of tracing patent claims, prosecution histories, and legal statuses across fragmented data sources.
Challenges in Patent Research
Patent research often requires more than retrieving a single document. Questions such as "What does claim 8 cover?" may involve examining the original application, issued claims, prosecution history, and subsequent legal proceedings. Traditional systems struggle with this complexity because patent information is scattered across various sources, including published patents, patent-office systems, and international databases. Each source has unique boundaries, limitations, and data formats, making it difficult to consolidate into a single index.
Harvey's Agentic Retrieval System
Harvey's patent research agent begins with a clear research objective, preserving the full context of the user's query, including patent numbers, claim details, and legal context. The system identifies the required information and selects eligible sources dynamically. For example, a query about claim scope may involve retrieving issued claims, examining reexamination records, and consulting PTAB (Patent Trial and Appeal Board) decisions. The agent prioritizes authoritative sources, such as the USPTO, and uses alternatives like Google Patents when necessary.
The system employs a routing table to guide source selection. Each source specifies the data it contains, its strengths, and fallback options. This ensures that the agent can adapt to incomplete or unavailable data by redirecting searches to alternative sources. The research path evolves as new information is discovered, allowing the agent to refine its focus and stop when the objective is either fully supported or the remaining gaps are clearly identified within the research budget.
Impact and Future Applications
Currently, users can integrate Harvey's Patent Research as a knowledge source, benefiting from its objective-driven, adaptive retrieval model. Over time, this approach could extend to non-patent literature, international patent-office records, and licensed or private datasets. By decoupling workflows from specific providers, the system offers a scalable framework for diverse legal and technical research tasks.
Harvey's agentic retrieval system represents a significant step forward in legal AI, providing a robust solution to the challenges of navigating fragmented and dynamic patent information landscapes.


