Query Decomposition

Decomposition identifies the logical dependencies between sub-questions, ensuring that foundational issues—such as establishing jurisdiction or a prima facie case—are resolved before addressing dependent elements like damages or affirmative defenses. This creates a directed acyclic graph (DAG) of reasoning steps.

• Linear Chains: A single path where each answer feeds the next query.
• Parallel Branches: Independent sub-questions (e.g., duty and causation) resolved concurrently.
• Convergent Synthesis: Multiple independent findings merge to satisfy a multi-prong legal test.

A core decomposition strategy splits a legal query into fact-finding sub-questions and legal-rule sub-questions. Factual queries retrieve evidence from the case record, while legal queries retrieve statutes and precedents. This separation prevents the model from conflating a party's allegation with established legal doctrine.

• Factual Sub-Query: "What was the defendant's blood alcohol content at the time of the accident?"
• Legal Sub-Query: "What is the statutory threshold for per se intoxication under Vehicle Code § 23152(b)?"
• Application Step: A final synthesis step applies the retrieved rule to the retrieved fact.

Complex legal tests are broken down into their constituent prima facie elements. A negligence query, for example, decomposes into four distinct sub-queries targeting duty, breach, causation, and damages. Each element retrieves its own supporting authority and factual evidence, enabling granular citation grounding.

• Duty: "Did the defendant owe a duty of care to the plaintiff under the common carrier doctrine?"
• Breach: "Did the defendant's failure to inspect the premises fall below the industry standard of care?"
• Causation: "Is there evidence that the breach was the proximate cause of the plaintiff's injury?"

Decomposition explicitly scopes sub-queries to a specific sovereign jurisdiction and temporal point-in-time. A broad question about 'liability for data breach' is decomposed into sub-queries that specify the governing law (e.g., CCPA vs. GDPR) and the effective date of the statute. This prevents cross-jurisdictional contamination in retrieval.

• Jurisdictional Filter: "Under California Civil Code § 1798.150, what constitutes a 'reasonable security procedure'?"
• Temporal Binding: "What was the text of the statute as amended effective January 1, 2023?"
• Hierarchy Awareness: Sub-queries can target binding authority (Supreme Court) before persuasive authority (sister circuits).

An advanced strategy generates sub-questions based on competing interpretive hypotheses. For an ambiguous contract clause, decomposition creates sub-queries that test each party's likely interpretation against canons of construction and extrinsic evidence. This mirrors adversarial legal reasoning.

• Hypothesis A: "Does the term 'consequential damages' in this MSA exclude lost profits under New York law?"
• Hypothesis B: "Does the carve-out in Section 12.3 explicitly include lost profits as direct damages?"
• Contra Proferentem Query: "Which party drafted the disputed clause, and how does that affect interpretation?"

Decomposition is not a single static step. An initial decomposition may produce sub-questions that are too broad, yielding noisy retrieval. A self-corrective loop evaluates the specificity of each sub-query and refines it—adding parties, timeframes, or legal terms of art—until the retrieval precision meets a confidence threshold.

• Initial Query: "What are the damages for breach?"
• Refined Query: "What is the measure of direct damages for breach of a SaaS subscription agreement under the UCC as adopted in Delaware?"
• Trigger: Refinement is triggered when initial retrieval returns documents from irrelevant practice areas or jurisdictions.

An iterative search process where the answer to an initial sub-query is used to formulate a secondary query to find connecting authority. This enables the construction of a logical evidence chain across multiple documents. Query decomposition is the planning phase that defines the sequence of hops; multi-hop retrieval is the execution phase that gathers the evidence for each decomposed step.

A reasoning framework where a language model explicitly generates a sequence of interconnected legal citations to demonstrate the logical derivation of a conclusion from primary authority. After decomposing a complex question, each sub-answer must be grounded in a verifiable citation chain. This technique ensures that the final synthesized response is not just plausible but forensically traceable to its sources.

A dynamic framework that routes a legal query to different processing paths based on its assessed complexity. Simple definitional questions bypass decomposition entirely, while multi-faceted jurisdictional analyses trigger full hierarchical decomposition. This prevents the system from over-engineering simple lookups or under-engineering complex reasoning tasks, optimizing both latency and accuracy.

A decoding strategy where the model alternates between generating a sentence of legal reasoning and issuing a new search query to gather additional evidence for the next reasoning step. Unlike static decomposition, this approach allows the system to adapt its sub-question plan on the fly based on information discovered mid-generation, creating a tight feedback loop between retrieval and synthesis.

Forward-Looking Active REtrieval monitors the model's generation confidence in real-time. When the model is about to generate a low-probability token—indicating uncertainty about a legal proposition—FLARE proactively searches for supporting authority. This serves as an implicit, confidence-driven form of query decomposition that triggers sub-queries only when the model recognizes a gap in its parametric knowledge.

A self-reflective architecture that evaluates the relevance of retrieved documents after each retrieval step. If the initial retrieval for a decomposed sub-question yields low-quality or irrelevant results, CRAG triggers a corrective action—such as a web search or knowledge graph lookup—before the flawed evidence contaminates the final synthesis. This ensures each atomic sub-answer is built on solid ground.