Adaptive RAG

A lightweight classification layer that analyzes the incoming legal query before retrieval begins. It assesses factors like:

• Entity count: Number of distinct legal entities mentioned
• Jurisdictional scope: Single vs. multi-jurisdictional questions
• Temporal depth: Whether the query involves historical statutory versions
• Logical operators: Presence of conditional or comparative language

The classifier outputs a complexity score that determines the processing path.

For low-complexity queries where the model's parametric knowledge is sufficient. Examples include:

• Defining a well-known legal term of art
• Explaining a foundational doctrine like res judicata
• Answering questions about widely codified procedural rules

This path bypasses retrieval entirely, minimizing latency to sub-second response times for simple lookups.

Triggered for moderate-complexity queries requiring one round of evidence gathering. The system:

1. Executes a single hybrid search (dense + sparse) against the legal corpus
2. Applies jurisdictional filtering to constrain results
3. Runs semantic re-ranking on the top candidates
4. Grounds the generation in the retrieved passages

Ideal for questions about a specific statute's text or a single case's holding.

Reserved for high-complexity queries demanding iterative reasoning. The system employs query decomposition to break the question into sub-questions, then:

• Performs chain-of-citation traversal through the authority graph
• Uses retrieval-interleaved generation to search after each reasoning step
• Applies temporal decay weighting to prioritize current precedent
• Synthesizes findings across multiple documents into a coherent answer

This path trades latency for citation-grounded accuracy on complex legal analysis.

Adaptive RAG incorporates Corrective RAG (CRAG) principles as a safety net. After initial retrieval, a lightweight evaluator scores the relevance of returned documents. If the score falls below a confidence threshold, the system:

• Triggers a reformulated query with expanded legal synonyms
• Falls back to knowledge graph traversal for deterministic entity relationships
• Escalates from single-hop to multi-hop if intermediate evidence is insufficient

This ensures that routing errors do not result in hallucinated legal answers.

The routing decision explicitly balances speed vs. thoroughness:

• Direct: < 500ms, suitable for real-time legal chat interfaces
• Single-hop: 1-3 seconds, appropriate for document Q&A
• Multi-hop: 5-15 seconds, justified for complex motion drafting support

A budget-conscious routing policy can cap multi-hop invocations or require explicit user confirmation before engaging deep reasoning paths, controlling infrastructure costs.

The technique of breaking a complex, multi-faceted legal question into a set of simpler sub-questions that can be answered independently. An Adaptive RAG router often uses a lightweight classifier to detect multi-part queries and triggers decomposition before retrieval.

• Enables parallel retrieval for distinct legal issues
• Reduces the cognitive load on the generator
• Example: Splitting 'Was the contract breached and what are the damages?' into two separate retrieval paths

An iterative search process where the answer to an initial query is used to formulate a secondary query to find connecting authority. Adaptive RAG routes complex, precedent-heavy questions to this path instead of a single retrieval step.

• Builds logical evidence chains across cases
• Essential for 'Why' and 'How' legal questions
• Example: Finding a statute, then retrieving the cases that interpreted its specific subsection

A post-retrieval step where a computationally intensive cross-encoder model re-orders a candidate list of legal documents. Adaptive RAG often applies this only to the high-complexity path to balance precision and latency.

• Processes query and passage simultaneously for deep relevance scoring
• Significantly improves precision@k for nuanced legal queries
• Trade-off: Higher latency than bi-encoder retrieval alone

A retrieval strategy that combines dense vector embeddings (semantic similarity) with sparse lexical scoring like BM25 (exact keyword matching). The Adaptive RAG router can weight these components differently based on whether the query is conceptual or citation-specific.

• Captures both meaning and precise legal terminology
• Essential for queries containing specific statute numbers
• Uses reciprocal rank fusion to merge result sets