Multi-Hop Legal Retrieval

The engine of multi-hop retrieval. The system does not stop after a single search. Instead, the initial query surfaces a primary document (e.g., a statute). The system then analyzes this output to extract new search terms—such as a case citation or a defined term of art—to formulate a secondary query.

• Bridge Entities: The specific legal concepts or citations that link Hop 1 to Hop 2.
• Expansion vs. Drift: The system must balance finding connecting authority with the risk of moving semantically away from the original legal question.
• Example: Query: 'Is a chatbot a contract agent?' → Hop 1 retrieves the Uniform Electronic Transactions Act → Hop 2 searches for '{UETA Section 2} agent definition court interpretation'.

The logical output of multi-hop retrieval is not just a list of documents, but a directed acyclic graph (DAG) of authority. Each node is a legal source, and each edge represents a logical relationship (e.g., 'interprets', 'applies', 'overrules').

• Citation Path Integrity: The system must verify that the path from the statute to the final case is legally valid and not based on overruled precedent.
• Synthesis: The final step involves merging the holdings from multiple hops into a single, non-contradictory statement of the law.
• Contrast: Unlike single-hop retrieval which returns a flat list, multi-hop constructs a hierarchical reasoning structure.

Before any search occurs, a complex legal question is often decomposed into a set of sub-questions that must be answered sequentially. This planning phase defines the retrieval strategy.

• Decomposition Logic: 'Is a chatbot liable for negligent misstatement?' decomposes into: (1) What is the legal definition of an 'agent'? (2) Does a chatbot satisfy that definition? (3) What are the elements of negligent misstatement?
• Dependency Mapping: The system identifies that Sub-Question 2 cannot be answered until Sub-Question 1 is resolved, creating a strict execution order.
• Parallelization: Independent sub-questions (e.g., defining 'agent' and defining 'misstatement') can be retrieved in parallel to reduce latency.

A critical memory mechanism that prevents information loss between hops. The specific text, citations, and metadata retrieved in Hop 1 are injected into the context of the query for Hop 2.

• State Management: The retriever maintains a running state of all previously gathered facts to avoid redundant searches.
• Query Enrichment: The Hop 2 query is not just the extracted entity; it is the original question plus the specific holding from Hop 1, ensuring contextual relevance.
• Example: Hop 2 query becomes: 'Regarding {Chatbot as Agent under UETA}, find cases discussing the application of the reasonable person standard to automated outputs.'

Multi-hop retrieval acts as a powerful hallucination mitigation technique. By forcing the model to explicitly link every logical step to a retrieved source, it prevents the generation of legally plausible but factually unsupported bridges.

• Verifiable Paths: Every connection between two legal concepts must be supported by a retrieved document, not the model's internal weights.
• Failure Mode: If the system cannot find a document linking Hop 1 to Hop 2, it must explicitly state the gap rather than invent a connection.
• Contrast with Single-Hop: A single retrieval might provide the correct statute but allow the model to hallucinate its application; multi-hop demands provenance for the application itself.

The decision logic that determines when to stop searching. The system evaluates the current evidence chain against a termination condition to prevent infinite loops or excessive latency.

• Termination Conditions: (1) A direct on-point case from a binding jurisdiction is found. (2) A statutory definition is reached that requires no further interpretation. (3) Maximum hop depth is reached.
• Confidence Scoring: A low-confidence score on the synthesized answer can trigger an additional retrieval hop to find supplementary authority.
• Adaptive Depth: Simple questions resolve in 1-2 hops; complex constitutional questions may require 4-5 hops to trace back to foundational precedent.