Databricks Document Intelligence: pushing the frontier for complex document extraction
Databricks introduces Precision Mode in its AI Extract document extraction API, improving accuracy on complex, long, and schema-heavy documents by combining custom-trained models with an agentic framework.
Databricks Document Intelligence processes billions of pages monthly for customers such as Panasonic, EY-Parthenon, and Intercontinental Exchange (NYSE), converting unstructured documents into structured data for enterprise workflows. The company identified persistent extraction challenges where existing large language model (LLM) or rules-based solutions underperform, particularly with long documents, large outputs, and deeply nested schemas. To address these gaps, Databricks developed Precision Mode within its ai_extract API, integrating custom-trained extraction models with an agentic harness designed for reliability and precision. The feature targets enterprise use cases where accuracy is critical, including financial filings, technical manuals, and government applications.
Precision Mode was evaluated against approximately 9,000 complex documents, including 10-K filings, bills of lading, and clinical notes, with benchmarks designed to stress-test existing extraction approaches. The evaluation covered documents up to 2,000 pages, invoices with thousands of line items, dense tables, and schemas with over 300 nested fields. Databricks compared Precision Mode against a chunk-and-merge baseline using leading models such as GPT, Claude, and Gemini, simulating real-world engineering practices where single-model calls fail to deliver complete results.
The results showed Precision Mode achieving 94.7% accuracy, outperforming the strongest frontier model baseline, GPT-5.6 Sol, by seven percentage points. On long-document workloads, frontier models frequently encountered operational failures such as timeouts, truncated outputs, and schema mismatches during merging. Precision Mode’s agentic approach mitigated these issues, maintaining efficiency and accuracy across complex extraction tasks. The evaluation defined accuracy as the fraction of correctly extracted objects, with scoring tailored to data types including primitives, strings, arrays, and nested objects.
Precision Mode is now available for customers requiring high-accuracy document extraction. Users can enable the feature by setting the mode to 'precision' when calling the ai_extract function or toggling it in the Information Extraction UI on the Agents page. The release follows Databricks’ broader effort to enhance document processing capabilities for enterprises handling large volumes of unstructured data.