Databricks Says Its New Extraction Mode Beats Frontier Models on the Documents That Break Them

Databricks is promising a more reliable way to turn the documents that overwhelm ordinary AI extraction into structured data. Its new Precision Mode for AI Extract combines task-specific models with an agentic system that breaks a job into smaller pieces, then rebuilds a single answer—an approach aimed at long files, giant outputs and schemas that require reasoning rather than simple lookup.

The company says Precision Mode reached 94.7% extraction accuracy across roughly 9,000 documents, ahead of its strongest tested frontier-model chunk-and-merge baseline, GPT-5.6 Sol, by seven percentage points. Those figures are Databricks’ own benchmark results, not an independent evaluation.

The failure is often in the workflow, not just the model

Databricks focuses on three tasks that get harder as documents and requested outputs grow. A lease may require matching a renewal clause on page 1 with a later definition. An invoice may list thousands of items. A financial classification may require combining several statements and applying one rule consistently.

• Long documents can require cross-page reconciliation of terms and references.
• Large, nested outputs can involve hundreds or thousands of extracted items.
• Reasoning-heavy schemas can require calculations or synthesis across a document.

The system’s answer is to customize both layers. Databricks trained document-extraction models around difficult customer-style workloads, then placed them inside a harness that semantically decomposes a large request, runs smaller extraction tasks in parallel, saves intermediate results and reconciles them into one structured output.

A tougher comparison than one giant prompt

Databricks did not compare Precision Mode only with a single request containing the entire document and schema. It used chunk-and-merge: split a document into smaller sections, extract each section independently, then combine the results. The company tested that pattern with leading GPT, Claude and Gemini models using their default API settings.

That is a consequential baseline because Databricks says one-call approaches can run into context limits on dense, lengthy documents. In its testing, frontier-model chunk-and-merge runs also encountered timeouts, truncated outputs and incomplete final merges that did not match the requested schema.

The benchmark is broad, but company-built in part

The evaluation included 10 internal datasets inspired by difficult customer workloads and five public benchmarks. It covered files as long as 2,000 pages, documents with dense tables and charts, invoices with thousands of line items, and schemas containing more than 300 deeply nested fields. The set spans financial, manufacturing, healthcare, legal and technical material.

Databricks defines accuracy as the share of extracted objects that match ground truth. The score uses direct matches for some basic values. Strings may also receive fuzzy matching or review by an LLM judge, while arrays are matched with their closest expected pairs. The aggregate score is useful, but it depends on the company’s benchmark design and scoring rules.

Available now for the hardest extraction jobs

Precision Mode is available in the ai_extract function and as a toggle in the Information Extraction interface on Databricks’ Agents page. The practical test is whether its staged extraction and reconciliation can preserve accuracy when production documents differ from the long files, large outputs and reasoning-heavy schemas used in the company’s evaluation.