Legal document review · Production

How Agentic Search Unlocks Legal Research Intelligence at Harvey

The problem

Traditional retrieval systems reduce legal research to a single query-and-retrieve operation, failing to handle the iterative, multi-source reasoning that complex legal work demands.

First attempt

The initial agentic prototype had three failure modes: the model queried only one or two sources when broader coverage was needed, treated nearly every query the same regardless of complexity, and the citation system was built and benchmarked only for single-source queries.

Workflow diagram · grounded in source
1
Query Understanding and Planning
ai_action
“The agent analyzes what information is needed and develops a search strategy. For complex queries, it identifies which knowledge sources will be relevant.”
2
Dynamic Tool Selection and Retrieval
ai_action
“The agent decides which sources to query and formulates specific search queries for each. It then executes searches and retrieves relevant information.”
3
Reasoning and Synthesis
ai_action
“The agent reasons about how the retrieved information connects and applies legal standards to specific facts or provisions.”
4
Completeness Check
validation
“The agent evaluates whether it has sufficient information to fully address the query. If gaps remain, it returns to Step 2 for additional retrieval rounds with refined queries.”
5
Citation-Backed Response
output
“the agent synthesizes findings into a comprehensive response with specific citations to source documents”
6
Expert Evaluation Feedback Loop
feedback_loop
“ALRs and SBDLs create evaluation queries that capture real usage patterns, spanning from straightforward searches to multi-jurisdictional analysis.”
Reported outcome

Harvey's privacy-preserving evaluation approach improved tool selection precision from near zero to 0.8-0.9, and complex queries now appropriately scale to 3-10 retrieval operations instead of a single tool call.

Reported metrics
Tool selection precision0.8-0.9
Retrieval operations for complex queries3-10
Legal knowledge sources150+
Reported stack
HarveyVaultLexisNexisiManageOpenAI Agent SDKLangSmith
Source
https://www.harvey.ai/blog/how-agentic-search-unlocks-legal-research-intelligence
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

Harvey's privacy-preserving evaluation approach improved tool selection precision from near zero to 0.8-0.9, and complex queries now appropriately scale to 3-10 retrieval operations instead of a single tool call.

What tools did this team use?

Harvey, Vault, LexisNexis, iManage, OpenAI Agent SDK, LangSmith.

What results were reported?

Tool selection precision: 0.8-0.9; Retrieval operations for complex queries: 3-10; Legal knowledge sources: 150+ (source-reported, not independently verified).

What failed first in this deployment?

The initial agentic prototype had three failure modes: the model queried only one or two sources when broader coverage was needed, treated nearly every query the same regardless of complexity, and the citation system…

How is this legal document review AI workflow structured?

Query Understanding and Planning → Dynamic Tool Selection and Retrieval → Reasoning and Synthesis → Completeness Check → Citation-Backed Response → Expert Evaluation Feedback Loop.