Corrective RAG (CRAG)

A lightweight classifier or confidence scoring module that assesses the relevance of each retrieved document chunk against the user's legal query. This evaluator acts as a gatekeeper, assigning a confidence score—high, ambiguous, or incorrect—to the initial retrieval set before generation begins. Documents scored as incorrect are discarded entirely, while ambiguous results trigger corrective action. This prevents the language model from hallucinating based on irrelevant or tangentially related case law.

When the retrieval evaluator flags results as ambiguous, CRAG triggers a structured lookup against a legal knowledge graph rather than relying solely on vector similarity. This fallback queries deterministic relationships—such as statutory hierarchies, citation networks, and doctrinal classifications—to inject grounded factual context into the generation prompt. For example, if a query about 'promissory estoppel elements' retrieves only dissenting opinions, the knowledge graph can surface the canonical Restatement section and its interpreting majority holdings.

For queries where the internal legal corpus lacks sufficient coverage, CRAG automatically formulates a corrective web search query. The system extracts key legal terms of art and canonical citations from the original query, executes a search against external legal databases or the open web, and integrates the retrieved content back into the context window. This ensures that even novel or rapidly evolving legal questions—such as recent regulatory changes—receive grounded, current responses rather than outdated or absent internal knowledge.

After corrective retrieval, CRAG applies a secondary filtering step to the combined document set—original high-confidence chunks, knowledge graph data, and web search results. This refinement process:

• De-duplicates redundant information across sources
• Re-ranks passages by composite relevance and authority scores
• Segments content into distinct knowledge strips for granular selection Only the most salient, non-redundant passages survive to populate the final generation context, preventing context window bloat and improving citation precision.

CRAG implements a generate-then-verify cycle where the model's output is scrutinized against the retrieved evidence. If generated claims lack direct support in the provided context, the system can trigger a secondary corrective retrieval pass—reformulating the query based on the model's own reasoning gaps. This recursive loop mirrors a legal associate's process of checking citations and filling logical holes, ensuring that every proposition in the final output is traceable to a specific, verifiable source document.

CRAG employs a dynamic routing mechanism that directs queries along different processing paths based on retrieval confidence:

• High confidence: Direct generation from retrieved documents without corrective action
• Ambiguous confidence: Knowledge graph augmentation before generation
• Low confidence: Full corrective web search pipeline This gating optimizes latency and compute cost, reserving expensive corrective operations only for queries where the initial retrieval quality is demonstrably insufficient. The routing decision is logged for observability and audit trails.