TechPrufer
Get started
Harvey
Harvey logoHarvey · Feature trail

US Case Law Knowledge Source

8 posts · Jul 2024 – Aug 2026

What this trail covers

Harvey's platform continues to evolve its AI-powered legal research capabilities, focusing on the integration and reasoning over authoritative legal sources. The system leverages advanced AI to move beyond traditional keyword-based searches, providing synthesized, reasoned answers grounded in primary and secondary legal content. This evolution emphasizes source transparency, verification, and the ability to handle complex legal questions efficiently. The platform aims to complement, rather than replace, traditional legal research methods by providing direct access to millions of US case law opinions, enabling faster research, analysis, and drafting within a single platform. It integrates with existing tools like the Word Add-In and Agent Builder for repeatable workflows.

All Harvey trails
  1. AI Tax Research: Get Reliable Answers to Complex QuestionsThis post details how AI tax research, specifically Harvey's platform, addresses complex, multi-jurisdictional tax questions by analyzing various authorities (IRC, Treasury Regulations, rulings, case law) and providing grounded, cited answers. It highlights the difference between general AI chatbots and purpose-built AI tax research in terms of knowledge sources, citations, currency, tax reasoning, and security. The post emphasizes that AI supports, but does not replace, professional judgment in verifying analysis, applying professional judgment, and defending conclusions. It also touches on the security bar for client tax data and best practices for using AI in tax research, including providing complete facts and using secure platforms. Harvey's approach is described as a secure, citation-based platform for researching complex tax questions, with a mention of co-building custom tax models with PwC.
  2. AI in Legal Research vs. Traditional Tools: What ChangedThis post details the fundamental shift in legal research from keyword-based searches to natural language questions, enabled by AI. It highlights the trade-offs between traditional methods (precision, control) and AI-native approaches (speed, synthesized answers). Key technical contributions discussed include the importance of source transparency, grounding AI answers in authoritative legal databases (not the open web), and the necessity of verifiable citations to combat hallucination. The post emphasizes that AI legal research platforms integrate existing research databases as grounding sources, changing how fast a question becomes a verified answer. It also introduces evaluation criteria for AI legal research platforms, focusing on source grounding, citation traceability, handling of negative treatment, performance measurement on realistic tasks, and data handling. Harvey's approach is presented as an example of agentic research over authoritative sources with grounded, verifiable answers.
  3. What to Know Before Adopting AI for Case Law ResearchThis post details the architectural considerations and evaluation criteria for AI case law research tools, emphasizing the importance of Retrieval-Augmented Generation (RAG) for grounding AI outputs in verified legal databases. It highlights key differentiators for reliable tools: surfacing adverse authority, understanding jurisdictional precision, tracking case treatment and history, providing transparent reasoning, and ensuring accurate pin cites. The post also discusses the impact of AI on associate training, shifting focus to verification and judgment, and the trend towards integrated AI platforms over standalone research tools.
  4. How to Use AI for Legal Questions ReliablyThis post details how Harvey's AI platform enhances reliability for legal questions by focusing on RAG, citation transparency, and precise question framing. It introduces the failure modes of fabricated citations, context loss, and confident-but-wrong statutory interpretation. The post emphasizes that the user (lawyer) always carries the verification burden and that Harvey's system facilitates this by surfacing sources and reasoning steps. It also outlines best practices for asking better legal questions, such as specifying jurisdiction, constraining the AI to defined documents, requesting citations, breaking down complex questions, and treating initial answers as drafts. The tiered approach for different user types (consumer, solo/small firm, enterprise) is also discussed, highlighting the increasing importance of security, privilege, and integration for enterprise users.
  5. US Case Law Content in HarveyThis post introduces the US Case Law knowledge source, which provides direct access to millions of opinions from CourtListener. It enables lawyers to conduct research, analysis, and drafting within Harvey, with responses including direct citations to the source. The capability allows searching across case law and uploaded files simultaneously and integrates with the Word Add-In for referencing case law during drafting. It can also be embedded in custom agents via Agent Builder for repeatable research workflows.
  7. Accelerating Legal Work With AI for Legal ResearchThis post details how AI legal research works, emphasizing its ability to go beyond keyword search to synthesize, compare, and cite legal authorities. It highlights the importance of interpreting natural-language questions, searching trusted legal sources (including internal firm knowledge and external databases), synthesizing and ranking results based on legal relevance and precedential weight, and returning citation-backed analysis for verification. The post also stresses the integration of research into analysis and drafting workflows, the need for jurisdiction-specific analysis, multi-source comparison, traceable citations, workflow integration, and permissions/governance controls. It positions AI as most effective for early-stage issue spotting, first-pass research, jurisdictional comparisons, summarizing authorities, analyzing cited authorities, and large-matter analysis.
  8. Introducing BigLaw Bench: ResearchIntroduces BLB: Research, a new benchmark dataset for evaluating foundation models on complex, agentic legal research tasks. This benchmark requires models to operate end-to-end, using tools to perform searches and provide grounded, cited responses. It aims to identify model failure modes in research and investigate improvements through agentic systems and access to non-public data. The benchmark includes realistic, difficult tasks across various practice areas, assessing how well models can synthesize information and provide helpful answers to legal professionals.
  9. Using Agents to Scale Harvey’s Knowledge SourcesThis post details the development of 'The Data Factory,' an automated pipeline designed to scale Harvey's knowledge source coverage from six to over 60 jurisdictions and from 20 to over 400 legal data sources. Key technical contributions include: an Intake Engine with a Sourcing Agent for automated discovery and validation of legal sources, and a Legal Review Agent that accelerates compliance review by pre-analyzing terms of use and extracting key clauses, doubling attorney review throughput. The system replaces hand-tuned configurations with a declarative config layer for defining jurisdictions, enabling parameterized tools for a single unified reasoning agent. The evaluation pipeline is a four-step process involving answer-first scenario generation, production simulation, trace validation, and multi-agent quality assessment (URL Classification, Citation Quality, Answer Quality, Presentation Agents), culminating in a Decision Agent for pass/fail determination. This automation significantly reduces the time and cost associated with integrating new legal data sources.
  10. Harvey partners with Voyage to build custom legal embeddingsThis post details the development of a custom embedding model, `voyage-law-2-harvey`, by fine-tuning Voyage AI's `voyage-law-2` model on over 20 billion tokens of US case law. The fine-tuning process involved both raw case law text and a dataset of exemplar questions with expert annotations. The custom model reduces irrelevant material returned by nearly 25% compared to leading off-the-shelf models, with 1/3 of the embedding dimensionality, leading to storage and latency benefits. This enhances the retrieval system's ability to identify relevant cases and passages.