LlamaIndex
Day 6 after signup
Extract v2.5 is officially out! 🚀
Every tier just got more accurate, and the price per page didn't change. Yes, really. Plus $3,000 in credits to test the switch!
Email copy
Extract v2.5 is officially out! ⚡️ We rebuilt the agent harness behind Extract and tuned it around the places where document extraction usually goes wrong. Every tier is now more accurate, and per-page pricing stays the same. Each tier now beats the one above it. On ExtractBench, Cost Effective went from 87.1 to 93.9, which puts it ahead of the old Agentic tier. Agentic went from 89.8 to 95.8, ahead of the old Agentic Plus. Agentic Plus went from 95.1 to 96.4. Our extraction agents are also SOTA in price-performance on ExtractBench, across a wide range of cost points. We are the best tool for document extraction across documents of any complexity. Where you'll notice it 🤔 Long lists: Models tend to stop early or lose count once records repeat for pages. v2.5 validates the full list against your schema, so you get everything back, not just the first few pages. Records that cross a page break: When a record starts on one page and finishes on the next, v2.5 keeps it together. Before, a grant schedule would lose its purpose and address at the break. Scanned forms: When a printed value has a reviewer's date or handwritten note on top of it, v2.5 separates the two and returns only the value your schema asks for. ❗️$3,000 in credits to test the switch ❗️ If you're using another OCR tool today, we'll give you $3,000 in credits to try v2.5 on your own documents. Reply to this email and we'll get you started! Try it out here! Check out the launch video ⬆️ You can also read our full technical breakdown here on our blog. Happy extracting! Advanced Citations, now on Agentic 👀 Citations used to be available only on Agentic Plus. Agentic has them now as well, and grounding accuracy improved on both (46.8 to 80.6 on Agentic). Each extracted value points to the exact place on the page it came from. Combined with confidence scores, that lets reviewers check flagged values in seconds instead of searching through the PDF. Also new! Native spreadsheet extraction lets agents read workbook cells directly instead of a flattened version. Schemas can now go up to 3,200 fields. Check out Jerry's tweet about the launch! 🚀 We're so excited for you try Extract v2.5 out. Please let us know any and all feedback by replying to this email :) LlamaIndex, 405 Howard Street, San Francisco, California 94105, United States Unsubscribe Manage preferences