databricks Databricks Blog ·

Databricks Introduces Precision Mode in AI Extract for Complex Document Processing

blogaidatabricksgaengineer
feature

Databricks has launched Precision Mode for its AI Extract API, designed to tackle challenging document extraction scenarios that typically cause existing LLM and rules-based solutions to fail. This new mode addresses issues with long documents requiring cross-page reconciliation, large outputs like thousands of line items, and complex schemas demanding intricate reasoning. By combining custom-tuned extraction models with an agentic harness, Precision Mode achieves state-of-the-art accuracy, outperforming other frontier models by seven points on benchmarks. The feature is now generally available for enterprises looking to convert unstructured document data into structured intelligence.

  • Precision Mode for Advanced Document Extraction in AI Extract
  • Achieving Frontier Accuracy on Challenging Benchmarks
Features (1)
  • Precision Mode for Advanced Document Extraction in AI Extract

    Databricks introduces Precision Mode in its `ai_extract` API to enhance accuracy for complex enterprise documents. This mode excels at handling long documents, generating large nested outputs, and processing intricate schemas that require deep reasoning and computation. It leverages custom finetuned extraction models alongside an agentic harness that reasons in stages, parallelizes sub-extractions, and merges results reliably.

Notes (1)
  • Achieving Frontier Accuracy on Challenging Benchmarks

    Precision Mode was evaluated against roughly 9,000 complex documents, including those up to 2,000 pages and schemas with over 300 nested fields. It achieved 94.7% accuracy, outperforming leading frontier models (e.g., GPT-5.6 Sol) by seven points. This robustness comes from its agentic approach and custom models, which are designed to overcome common failure modes like context limits, truncated outputs, and incomplete merges seen in other solutions.

Read the original announcement →

https://www.databricks.com/blog/databricks-document-intelligence-pushing-frontier-complex-document-extraction

Related releases