Improving Enterprise Accuracy with Structured RAG

Traditional embedder-based RAG is strong at retrieving similar document fragments, but it often fails on queries that require aggregation, comparison, or exhaustive verification. In domains like finance, compliance, and regulatory documents—where missing even a single item is risky—this limitation can be critical.

AI21 Maestro's Structured RAG (S-RAG) converts documents into a pre-structured relational DB, and performs filtering, aggregation, and comparison via SQL at query time. For documents with repeating patterns, such as financial disclosures, resumes, and contracts, schemas can be automatically inferred or manually defined by users, and links to the original text are preserved to ensure traceability and control.

• For aggregation-type questions, it accurately handles max/min, averages, and trend analysis.
• For exhaustive-type questions, it finds items without omission when the schema is correct.
• In dense corpora, it first narrows down using precise attributes, then adds text search only when needed.

According to the article, this hybrid structure showed up to 60% higher accuracy on aggregation-type benchmarks, and achieved near-perfect recall on exhaustive-type questions. As a result, AI21 Maestro is presented as a more reliable retrieval-and-response engine for enterprise workloads involving large-scale documents, complex queries, and where auditability matters.