Databricks Launches AI Extract with 95% Accuracy and 500-Page Document Support

Databricks launched AI Extract around August 20, targeting agent workflows that extract fields from PDFs and other documents. Because LLMs trained on next-token prediction tend to "auto-correct" document content and distort data, AI Extract uses pretraining optimizations to reach 95% extraction accuracy at very low cost, outperforming other solutions at roughly 87%, and is usable directly via SQL.

Confirmed

• Handles documents exceeding 1 million tokens and 500+ pages, plus large-scale outputs such as invoices with thousands of line items
• Supports large nested schemas with 1000+ objects and complex schemas requiring frontier reasoning
• A Precision Mode combines custom models with an agent architecture, achieving 94.7% precision on 9000 documents
• Implementation combines custom models and agent architecture; accessible directly through SQL

Why it matters

Document extraction is a high-frequency enterprise need for agent deployment, and traditional LLM approaches fail precision requirements due to hallucinated "corrections." AI Extract breaks through on both ultra-long documents and complex nested structures, delivered at low cost via SQL, potentially lowering the barrier to building document agents.

2026-08-20 ~ 2026-08-21 · 5 related posts

• [source] Databricks releases AI Extract for processing 500+ page documents and complex nested schemas — matei_zaharia · 2026-08-20
• Databricks AI Extract achieves 95% accuracy on PDF field extraction — jefrankle · 2026-08-20
• [source] Databricks' AI Extract reaches 94.7% accuracy on 9,000 complex docs — mrdrozdov · 2026-08-21

2 near-duplicate retellings: Zachly · Scobleizer